{
  "schema_version": 1,
  "columns": [
    "configuration",
    "axis",
    "stratum",
    "metric",
    "events",
    "species",
    "source_groups",
    "method",
    "estimate",
    "ci_low",
    "ci_high",
    "valid_resamples",
    "interval_status"
  ],
  "rows": [
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.43356643356643354",
      "ci_low": "0.38564615681848224",
      "ci_high": "0.4819290676218387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.43356643356643354",
      "ci_low": "0.38564615681848224",
      "ci_high": "0.4819290676218387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5664335664335665",
      "ci_low": "0.5180709323781613",
      "ci_high": "0.6143538431815176",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.22525144798559688",
      "ci_low": "0.18581102096437715",
      "ci_high": "0.2327245798078507",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4650352474541355",
      "ci_low": "0.3836098497329077",
      "ci_high": "0.48046364863556285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3767313019390582",
      "ci_low": "0.3242506811989101",
      "ci_high": "0.429730288139379",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3767313019390582",
      "ci_low": "0.3242506811989101",
      "ci_high": "0.429730288139379",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.6232686980609419",
      "ci_low": "0.5702697118606209",
      "ci_high": "0.6757493188010899",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.19503914425228208",
      "ci_low": "0.1471057016719108",
      "ci_high": "0.20126324698524756",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3768552956739009",
      "ci_low": "0.28423813543386156",
      "ci_high": "0.3888815280731902",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5706666666666667",
      "ci_low": "0.5190208791964778",
      "ci_high": "0.6215506153351457",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5706666666666667",
      "ci_low": "0.5190208791964778",
      "ci_high": "0.6215506153351457",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.42933333333333334",
      "ci_low": "0.3784493846648543",
      "ci_high": "0.4809791208035223",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3569681825862559",
      "ci_low": "0.3021033359601287",
      "ci_high": "0.36228625823980637",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.576640910331644",
      "ci_low": "0.48801308116636166",
      "ci_high": "0.5852316479258409",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.15517241379310345",
      "ci_low": "0.10285328185328185",
      "ci_high": "0.21142857142857144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.15517241379310345",
      "ci_low": "0.10285328185328185",
      "ci_high": "0.21142857142857144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.8448275862068966",
      "ci_low": "0.7885714285714286",
      "ci_high": "0.8971467181467181",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.07969720820960491",
      "ci_low": "0.04284947343218079",
      "ci_high": "0.08733185199203032",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.12523847004366484",
      "ci_low": "0.0673348868219984",
      "ci_high": "0.13723576741604762",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.2727272727272727",
      "ci_high": "0.3955579710144927",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.2727272727272727",
      "ci_high": "0.3955579710144927",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.6044420289855073",
      "ci_high": "0.7272727272727273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.19571272127354958",
      "ci_low": "0.13889177057499402",
      "ci_high": "0.20008868122703138",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3022398986756082",
      "ci_low": "0.21449108873606673",
      "ci_high": "0.3089977102493396",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.3875",
      "ci_low": "0.33886158349595763",
      "ci_high": "0.43596059113300495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.3875",
      "ci_low": "0.33886158349595763",
      "ci_high": "0.43596059113300495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6125",
      "ci_low": "0.5640394088669951",
      "ci_high": "0.6611384165040424",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.15161317024192555",
      "ci_low": "0.13137522302392565",
      "ci_high": "0.16129822046561862",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4889524740302099",
      "ci_low": "0.4236850942521602",
      "ci_high": "0.52018676100162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4282051282051282",
      "ci_low": "0.3753351206434316",
      "ci_high": "0.48052038239538236",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4282051282051282",
      "ci_low": "0.3753351206434316",
      "ci_high": "0.48052038239538236",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5717948717948718",
      "ci_low": "0.5194796176046176",
      "ci_high": "0.6246648793565683",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.21273344644357361",
      "ci_low": "0.18372988850400115",
      "ci_high": "0.22799181345311462",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.45274041166196427",
      "ci_low": "0.39101489091877173",
      "ci_high": "0.4852133465797054",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4186046511627907",
      "ci_low": "0.36236670830320167",
      "ci_high": "0.4757325418177564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4186046511627907",
      "ci_low": "0.36236670830320167",
      "ci_high": "0.4757325418177564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5813953488372093",
      "ci_low": "0.5242674581822435",
      "ci_high": "0.6376332916967983",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2434829250774315",
      "ci_low": "0.19752116645740875",
      "ci_high": "0.25837787415102914",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4504434113932483",
      "ci_low": "0.3654141579462061",
      "ci_high": "0.47799906717940405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.337037037037037",
      "ci_low": "0.2753597939706105",
      "ci_high": "0.40206275683079806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.337037037037037",
      "ci_low": "0.2753597939706105",
      "ci_high": "0.40206275683079806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.662962962962963",
      "ci_low": "0.5979372431692019",
      "ci_high": "0.7246402060293895",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.14710353198555656",
      "ci_low": "0.11964032139184026",
      "ci_high": "0.16342022437109335",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3595864115202494",
      "ci_low": "0.29245411895783174",
      "ci_high": "0.3994716595737837",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4794520547945205",
      "ci_low": "0.411214953271028",
      "ci_high": "0.5483870967741935",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4794520547945205",
      "ci_low": "0.411214953271028",
      "ci_high": "0.5483870967741935",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5205479452054794",
      "ci_low": "0.45161290322580644",
      "ci_high": "0.5887850467289719",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.23276873172995055",
      "ci_low": "0.18646270087168929",
      "ci_high": "0.2479643557784681",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5360734427720074",
      "ci_low": "0.42942925049237535",
      "ci_high": "0.5710694254291993",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.410958904109589",
      "ci_low": "0.35472673471343685",
      "ci_high": "0.4686375691881918",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.410958904109589",
      "ci_low": "0.35472673471343685",
      "ci_high": "0.4686375691881918",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.589041095890411",
      "ci_low": "0.5313624308118081",
      "ci_high": "0.6452732652865631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.24694209428738365",
      "ci_low": "0.1948372711577632",
      "ci_high": "0.25961043333057826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.45172334320862867",
      "ci_low": "0.3564096423617619",
      "ci_high": "0.4748971341413017",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5052631578947369",
      "ci_low": "0.4293193717277487",
      "ci_high": "0.580808836871711",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5052631578947369",
      "ci_low": "0.4293193717277487",
      "ci_high": "0.580808836871711",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.49473684210526314",
      "ci_low": "0.41919116312828886",
      "ci_high": "0.5706806282722513",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.23259088134088132",
      "ci_low": "0.17902836413882756",
      "ci_high": "0.24230492691305333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.453835866030988",
      "ci_low": "0.3493236373440538",
      "ci_high": "0.4727901012937626",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.25842696629213485",
      "ci_low": "0.16091954022988506",
      "ci_high": "0.36170719351570413",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.25842696629213485",
      "ci_low": "0.16091954022988506",
      "ci_high": "0.36170719351570413",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.7415730337078652",
      "ci_low": "0.6382928064842959",
      "ci_high": "0.8390804597701149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.14100038850038848",
      "ci_low": "0.06028732816890711",
      "ci_high": "0.1448480861244019",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3845465140919686",
      "ci_low": "0.16441998591520124",
      "ci_high": "0.39504023488473244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3",
      "ci_low": "0.06666666666666667",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3681592039800995",
      "ci_low": "0.30041845493562236",
      "ci_high": "0.43783783783783786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3681592039800995",
      "ci_low": "0.30041845493562236",
      "ci_high": "0.43783783783783786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6318407960199005",
      "ci_low": "0.5621621621621622",
      "ci_high": "0.6995815450643775",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.1824968123362414",
      "ci_low": "0.13407583269900814",
      "ci_high": "0.1988665050943983",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3987893306606757",
      "ci_low": "0.29298052330524005",
      "ci_high": "0.43456014076183325",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2714285714285714",
      "ci_low": "0.16",
      "ci_high": "0.3888888888888889",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2714285714285714",
      "ci_low": "0.16",
      "ci_high": "0.3888888888888889",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.7285714285714285",
      "ci_low": "0.6111111111111112",
      "ci_high": "0.84",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.07522995022995023",
      "ci_low": "0.04524093249093249",
      "ci_high": "0.09490935609937158",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.39764402264402265",
      "ci_low": "0.23913064316635743",
      "ci_high": "0.5016637393823926",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.2826086956521739",
      "ci_high": "0.5714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.2826086956521739",
      "ci_high": "0.5714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5777777777777777",
      "ci_low": "0.42857142857142855",
      "ci_high": "0.717391304347826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.24641311069882502",
      "ci_low": "0.12083835807050093",
      "ci_high": "0.26335171568627447",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5307359307359308",
      "ci_low": "0.2602672327672328",
      "ci_high": "0.5672190799396681",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.32",
      "ci_low": "0.25225225225225223",
      "ci_high": "0.3914898635086311",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.32",
      "ci_low": "0.25225225225225223",
      "ci_high": "0.3914898635086311",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.68",
      "ci_low": "0.6085101364913689",
      "ci_high": "0.7477477477477478",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.1562515792924868",
      "ci_low": "0.12243740145269388",
      "ci_high": "0.17466069520351027",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3530128272904331",
      "ci_low": "0.27661783291164166",
      "ci_high": "0.3946037928671899",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5633802816901409",
      "ci_low": "0.44943413124898224",
      "ci_high": "0.6764705882352942",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5633802816901409",
      "ci_low": "0.44943413124898224",
      "ci_high": "0.6764705882352942",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.43661971830985913",
      "ci_low": "0.3235294117647059",
      "ci_high": "0.5505658687510177",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3815721315721316",
      "ci_low": "0.24297884751326046",
      "ci_high": "0.39288137788137784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.55115974560419",
      "ci_low": "0.35096944640804284",
      "ci_high": "0.5674953236064346",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4186046511627907",
      "ci_low": "0.36236670830320167",
      "ci_high": "0.4757325418177564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4186046511627907",
      "ci_low": "0.36236670830320167",
      "ci_high": "0.4757325418177564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5813953488372093",
      "ci_low": "0.5242674581822435",
      "ci_high": "0.6376332916967983",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2434829250774315",
      "ci_low": "0.19752116645740875",
      "ci_high": "0.25837787415102914",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4504434113932483",
      "ci_low": "0.3654141579462061",
      "ci_high": "0.47799906717940405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5897435897435898",
      "ci_low": "0.474570838765754",
      "ci_high": "0.7031292925824175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5897435897435898",
      "ci_low": "0.474570838765754",
      "ci_high": "0.7031292925824175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.41025641025641024",
      "ci_low": "0.2968707074175824",
      "ci_high": "0.525429161234246",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.30847118889882047",
      "ci_low": "0.2236420415521978",
      "ci_high": "0.33856595610272083",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.616942377797641",
      "ci_low": "0.4472840831043956",
      "ci_high": "0.6771319122054416",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.41975308641975306",
      "ci_low": "0.36569402305825244",
      "ci_high": "0.4741648691048142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.41975308641975306",
      "ci_low": "0.36569402305825244",
      "ci_high": "0.4741648691048142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5802469135802469",
      "ci_low": "0.5258351308951859",
      "ci_high": "0.6343059769417475",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.1783428019138341",
      "ci_low": "0.15185341092813434",
      "ci_high": "0.1896237204681453",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.49748255270701086",
      "ci_low": "0.42359109364163783",
      "ci_high": "0.5289503781479845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.15",
      "ci_high": "0.3563218390804598",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.15",
      "ci_high": "0.3563218390804598",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.75",
      "ci_low": "0.6436781609195402",
      "ci_high": "0.85",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.12016272418058131",
      "ci_low": "0.05414349489795918",
      "ci_high": "0.13376677713246035",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.33645562770562776",
      "ci_low": "0.1516017857142857",
      "ci_high": "0.37454697597088904",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.47619047619047616",
      "ci_low": "0.3734820233314209",
      "ci_high": "0.5798319327731093",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.47619047619047616",
      "ci_low": "0.3734820233314209",
      "ci_high": "0.5798319327731093",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5238095238095238",
      "ci_low": "0.42016806722689076",
      "ci_high": "0.6265179766685789",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.29517698525795694",
      "ci_low": "0.1912797510223981",
      "ci_high": "0.3149832404437667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.47966260104417996",
      "ci_low": "0.31082959541139693",
      "ci_high": "0.511847765721121",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4105263157894737",
      "ci_low": "0.35116993030202454",
      "ci_high": "0.4709946780817955",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4105263157894737",
      "ci_low": "0.35116993030202454",
      "ci_high": "0.4709946780817955",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5894736842105263",
      "ci_low": "0.5290053219182044",
      "ci_high": "0.6488300696979754",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.22266362223470063",
      "ci_low": "0.17962296574783346",
      "ci_high": "0.23753944284421868",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.41107130258713964",
      "ci_low": "0.3316116290729233",
      "ci_high": "0.43853435602009605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.5918367346938775",
      "ci_low": "0.529143220735149",
      "ci_high": "0.6536964980544747",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.5918367346938775",
      "ci_low": "0.529143220735149",
      "ci_high": "0.6536964980544747",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.40816326530612246",
      "ci_low": "0.3463035019455253",
      "ci_high": "0.470856779264851",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.3709912983692708",
      "ci_low": "0.29522804535874536",
      "ci_high": "0.37082372741107344",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.5910708821476517",
      "ci_low": "0.47036332650376383",
      "ci_high": "0.5908039046888289",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5307692307692308",
      "ci_low": "0.4428544165757906",
      "ci_high": "0.6194029850746269",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5307692307692308",
      "ci_low": "0.4428544165757906",
      "ci_high": "0.6194029850746269",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.46923076923076923",
      "ci_low": "0.3805970149253731",
      "ci_high": "0.5571455834242094",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.3512735981966752",
      "ci_low": "0.24627254281666044",
      "ci_high": "0.35753108003108003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5189269064269064",
      "ci_low": "0.3638117109791575",
      "ci_high": "0.5281709136822773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.07894736842105263",
      "ci_low": "0.02531645569620253",
      "ci_high": "0.14492753623188406",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.07894736842105263",
      "ci_low": "0.02531645569620253",
      "ci_high": "0.14492753623188406",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.9210526315789473",
      "ci_low": "0.855072463768116",
      "ci_high": "0.9746835443037974",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.048546691403834254",
      "ci_low": "0.011544011544011544",
      "ci_high": "0.0567099567099567",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.08126293995859213",
      "ci_low": "0.01932367149758454",
      "ci_high": "0.09492753623188407",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.13541666666666666",
      "ci_high": "0.3",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.13541666666666666",
      "ci_high": "0.3",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.7857142857142857",
      "ci_low": "0.7",
      "ci_high": "0.8645833333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.1114969135802469",
      "ci_low": "0.058595792483660133",
      "ci_high": "0.12577215608465608",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.1745169082125604",
      "ci_low": "0.09171515345268542",
      "ci_high": "0.19686076604554864",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.3177570093457944",
      "ci_low": "0.23075221238938057",
      "ci_high": "0.4056673827760134",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.3177570093457944",
      "ci_low": "0.23075221238938057",
      "ci_high": "0.4056673827760134",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6822429906542056",
      "ci_low": "0.5943326172239866",
      "ci_high": "0.7692477876106193",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.19332045732045733",
      "ci_low": "0.11640574795574796",
      "ci_high": "0.2008696803196803",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2842947901771431",
      "ci_low": "0.1711849234643352",
      "ci_high": "0.2953965887054122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3458646616541353",
      "ci_low": "0.26153846153846155",
      "ci_high": "0.4296875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3458646616541353",
      "ci_low": "0.26153846153846155",
      "ci_high": "0.4296875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6541353383458647",
      "ci_low": "0.5703125",
      "ci_high": "0.7384615384615385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.20422734764840028",
      "ci_low": "0.1273655459317224",
      "ci_high": "0.20972793816543817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.30842497399962493",
      "ci_low": "0.19234796732545828",
      "ci_high": "0.3167319882498454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.20279720279720279",
      "ci_low": "0.16173021566427545",
      "ci_high": "0.24489910570969958",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.3613053613053613",
      "ci_low": "0.31279568657744744",
      "ci_high": "0.41067285382830626",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5612903225806452",
      "ci_low": "0.47770700636942676",
      "ci_high": "0.6456733923884513",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.1585081585081585",
      "ci_low": "0.12284904521746627",
      "ci_high": "0.19529499188210314",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.19813519813519814",
      "ci_low": "0.1592479431708823",
      "ci_high": "0.2375588235294117",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "30.62937062937063",
      "ci_low": "26.711023751023752",
      "ci_high": "34.661018338808276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "10.815850815850816",
      "ci_low": "5.569000091803129",
      "ci_high": "16.341476612818074",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.24428938270401684",
      "ci_low": "0.18414097253697911",
      "ci_high": "0.2530281419835335",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.3230924093827319",
      "ci_low": "0.2435412862585853",
      "ci_high": "0.3346501232685443",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3628808864265928",
      "ci_low": "0.3103448275862069",
      "ci_high": "0.41690705003501666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.631578947368421",
      "ci_low": "0.5782122905027933",
      "ci_high": "0.6840695488721804",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5745614035087719",
      "ci_low": "0.5064935064935064",
      "ci_high": "0.6431785765869863",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.26869806094182824",
      "ci_low": "0.22126436781609196",
      "ci_high": "0.3172062048798318",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3545706371191136",
      "ci_low": "0.3019929208987381",
      "ci_high": "0.4069767441860465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "50.05540166204987",
      "ci_low": "45.67032267957495",
      "ci_high": "54.48827238756173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "14.598337950138504",
      "ci_low": "6.922774869109948",
      "ci_high": "22.44656713164777",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.25608055726816586",
      "ci_low": "0.21684344650954213",
      "ci_high": "0.2680882591219495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.35590857111846774",
      "ci_low": "0.3013756375217365",
      "ci_high": "0.37259724149152307",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3253333333333333",
      "ci_low": "0.2761394101876676",
      "ci_high": "0.3763164968873797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.512",
      "ci_low": "0.4576247968423497",
      "ci_high": "0.5646437994722955",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6354166666666666",
      "ci_low": "0.564591372125328",
      "ci_high": "0.7055214723926381",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.18666666666666668",
      "ci_low": "0.14758165436115628",
      "ci_high": "0.2281556816627923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.22933333333333333",
      "ci_low": "0.18591420249244156",
      "ci_high": "0.27520435967302453",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "44.45333333333333",
      "ci_low": "39.950223501780435",
      "ci_high": "48.94329896907217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "21.520000000000003",
      "ci_low": "15.067376411842494",
      "ci_high": "28.01635932178827",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.2765855604090898",
      "ci_low": "0.22849024069612303",
      "ci_high": "0.28288112509819346",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3616888097657328",
      "ci_low": "0.298794930141084",
      "ci_high": "0.36992147128225294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.22413793103448276",
      "ci_low": "0.15950920245398773",
      "ci_high": "0.29081864564007415",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.39080459770114945",
      "ci_low": "0.3149171270718232",
      "ci_high": "0.4662585303646515",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5735294117647058",
      "ci_low": "0.44926142697881827",
      "ci_high": "0.691381766381766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.16666666666666666",
      "ci_low": "0.11299435028248588",
      "ci_high": "0.2236024844720497",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.25287356321839083",
      "ci_low": "0.18888364779874214",
      "ci_high": "0.31976744186046513",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "31.954022988505745",
      "ci_low": "25.91941757779647",
      "ci_high": "38.08755804036767",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "6.666666666666666",
      "ci_low": "-2.9069973610423885",
      "ci_high": "15.934285714285709",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.18976279650436953",
      "ci_low": "0.1136609270878934",
      "ci_high": "0.18685092593387373",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.21933621933621933",
      "ci_low": "0.13137431832237026",
      "ci_high": "0.21597055075473715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2375",
      "ci_low": "0.18375850340136055",
      "ci_high": "0.29338842975206614",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5125",
      "ci_low": "0.44680633065922565",
      "ci_high": "0.5757575757575758",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.4634146341463415",
      "ci_low": "0.37189534118335404",
      "ci_high": "0.5564533660589059",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.275",
      "ci_low": "0.2155154892066162",
      "ci_high": "0.3359691699604743",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3458333333333333",
      "ci_low": "0.2833315450643777",
      "ci_high": "0.4085636948062933",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "37.375",
      "ci_low": "32.38683127572016",
      "ci_high": "42.46754628350374",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "2.791666666666667",
      "ci_low": "-5.358803715977617",
      "ci_high": "11.153925619834709",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.18365873608592054",
      "ci_low": "0.13449501550229706",
      "ci_high": "0.19284751647425247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2394537951499977",
      "ci_low": "0.17535426071818475",
      "ci_high": "0.2514341037575697",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.12",
      "ci_low": "0.08644859813084112",
      "ci_high": "0.15625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2675",
      "ci_low": "0.2210796915167095",
      "ci_high": "0.31368040140584535",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4485981308411215",
      "ci_low": "0.3461448598130841",
      "ci_high": "0.5545454545454546",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1475",
      "ci_low": "0.11056378603548415",
      "ci_high": "0.18518518518518517",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.205",
      "ci_low": "0.1628498727735369",
      "ci_high": "0.24880382775119617",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "22.775",
      "ci_low": "19.331369249848212",
      "ci_high": "26.381354294068856",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "2.275",
      "ci_low": "-2.282142857142857",
      "ci_high": "7.0213857064920875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.10598496391563618",
      "ci_low": "0.08181920280091934",
      "ci_high": "0.12321767899743628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.14837894948189065",
      "ci_low": "0.11454688392128708",
      "ci_high": "0.1725047505964108",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4358974358974359",
      "ci_low": "0.3842239185750636",
      "ci_high": "0.48894365862597905",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.7076923076923077",
      "ci_low": "0.6601458407736052",
      "ci_high": "0.753695500330858",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.6159420289855072",
      "ci_low": "0.5546875",
      "ci_high": "0.6772911898505269",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.2717948717948718",
      "ci_low": "0.22538860103626943",
      "ci_high": "0.318769888140068",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.33589743589743587",
      "ci_low": "0.28690591868022713",
      "ci_high": "0.38536646802236285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "56.66666666666667",
      "ci_low": "52.44723819510215",
      "ci_high": "60.8857171474359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "23.076923076923073",
      "ci_low": "15.254609312983957",
      "ci_high": "31.041450402698416",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.2985159203125846",
      "ci_low": "0.2703997825889327",
      "ci_high": "0.3164523007682581",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4745637707533396",
      "ci_low": "0.429866321038816",
      "ci_high": "0.5030780166059488",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.15614617940199335",
      "ci_low": "0.116363350485992",
      "ci_high": "0.19931271477663232",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.38870431893687707",
      "ci_low": "0.3354222329394743",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4017094017094017",
      "ci_low": "0.31297709923664124",
      "ci_high": "0.491670081967213",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.23255813953488372",
      "ci_low": "0.18688439503091497",
      "ci_high": "0.28135845773936463",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2823920265780731",
      "ci_low": "0.2326388888888889",
      "ci_high": "0.33440514469453375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "27.70764119601329",
      "ci_low": "23.846091205211728",
      "ci_high": "31.74499307552999",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "-0.53156146179402",
      "ci_low": "-6.974614064329225",
      "ci_high": "5.94064150824028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.10846636868463494",
      "ci_low": "0.08329116758465686",
      "ci_high": "0.1246453781512605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16269955302695238",
      "ci_low": "0.12493675137698529",
      "ci_high": "0.18696806722689074",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.5343511450381679",
      "ci_high": "0.669016005121639",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.8111111111111111",
      "ci_low": "0.7591457851166103",
      "ci_high": "0.860264892477548",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.7397260273972602",
      "ci_low": "0.6712313340737999",
      "ci_high": "0.8075117370892019",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2111111111111111",
      "ci_low": "0.15476078167115903",
      "ci_high": "0.2684842089342921",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3111111111111111",
      "ci_low": "0.25",
      "ci_high": "0.37269538017960824",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "72.25925925925927",
      "ci_low": "67.6204114576178",
      "ci_high": "77.10209597352454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "41.148148148148145",
      "ci_low": "31.533094687046976",
      "ci_high": "51.08478797078075",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.4609452688603739",
      "ci_low": "0.4171875950328465",
      "ci_high": "0.49025018468780696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6487377858034892",
      "ci_low": "0.5871529115277099",
      "ci_high": "0.689981741412469",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.0410958904109589",
      "ci_low": "0.017241379310344827",
      "ci_high": "0.0700023364485981",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.2146118721461187",
      "ci_low": "0.1596244131455399",
      "ci_high": "0.2715578028293545",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.19148936170212766",
      "ci_low": "0.08333333333333333",
      "ci_high": "0.3148148148148148",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1735159817351598",
      "ci_low": "0.12389279045222051",
      "ci_high": "0.22667010309278343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.2009132420091324",
      "ci_low": "0.14601769911504425",
      "ci_high": "0.25870646766169153",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "15.342465753424657",
      "ci_low": "12.252223002223001",
      "ci_high": "18.585026608611514",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "-4.748858447488584",
      "ci_low": "-10.096618357487923",
      "ci_high": "0.4092803030303002",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.07278911564625849",
      "ci_low": "0.026239067055393583",
      "ci_high": "0.08969873663751214",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.10808080808080807",
      "ci_low": "0.03896103896103896",
      "ci_high": "0.1331890331890332",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.16095890410958905",
      "ci_low": "0.11960132890365449",
      "ci_high": "0.2048611111111111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.3972602739726027",
      "ci_low": "0.34275618374558303",
      "ci_high": "0.4539249146757679",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.4051724137931034",
      "ci_low": "0.3148148148148148",
      "ci_high": "0.4961832061068702",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.2363013698630137",
      "ci_low": "0.18892508143322476",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.2842465753424658",
      "ci_low": "0.2332128617133034",
      "ci_high": "0.3366675084175084",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "28.28767123287671",
      "ci_low": "24.39022062930553",
      "ci_high": "32.41499452129845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "-0.136986301369863",
      "ci_low": "-6.724738675958188",
      "ci_high": "6.448386533084807",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.12341732520131748",
      "ci_low": "0.08699156148951231",
      "ci_high": "0.13730731558166465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.18362089847025284",
      "ci_low": "0.12942646953317685",
      "ci_high": "0.20428649391418402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.11052631578947368",
      "ci_low": "0.06451612903225806",
      "ci_high": "0.16326530612244897",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.30526315789473685",
      "ci_low": "0.23976608187134502",
      "ci_high": "0.3742698226011957",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.3620689655172414",
      "ci_low": "0.22855889724310777",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.19473684210526315",
      "ci_low": "0.13829787234042554",
      "ci_high": "0.25500255102040814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.24736842105263157",
      "ci_low": "0.1827956989247312",
      "ci_high": "0.31382978723404253",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "23.105263157894736",
      "ci_low": "18.358588957055215",
      "ci_high": "28.243921899325375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "-1.6315789473684212",
      "ci_low": "-9.356764236795197",
      "ci_high": "6.300052083333332",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.14207650273224043",
      "ci_low": "0.07262507983819459",
      "ci_high": "0.1494331078757308",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.2113821138211382",
      "ci_low": "0.10805194805194807",
      "ci_high": "0.22232730683950197",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.1348314606741573",
      "ci_low": "0.0625",
      "ci_high": "0.2159090909090909",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3707865168539326",
      "ci_low": "0.25640243902439025",
      "ci_high": "0.48387945670628174",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.36363636363636365",
      "ci_low": "0.19230769230769232",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.23595505617977527",
      "ci_low": "0.1392317158931083",
      "ci_high": "0.33963207547169794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.29213483146067415",
      "ci_low": "0.189873417721519",
      "ci_high": "0.4",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "30.0",
      "ci_low": "22.345478595478596",
      "ci_high": "37.922196705733285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.7865168539325842",
      "ci_low": "-10.0",
      "ci_high": "12.222222222222223",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.28620689655172415",
      "ci_low": "0.0999904214559387",
      "ci_high": "0.2906444991789818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3772727272727273",
      "ci_low": "0.13180555555555556",
      "ci_high": "0.38312229437229434",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.125",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2222222222222222",
      "ci_high": "0.7777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.8333333333333334",
      "ci_low": "0.5",
      "ci_high": "1.0",
      "valid_resamples": "9978",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.23076923076923078",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.16666666666666666",
      "ci_low": "0.0",
      "ci_high": "0.375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "53.33333333333333",
      "ci_low": "28.0",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "36.666666666666664",
      "ci_low": "4.0",
      "ci_high": "74.54545454545455",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.35",
      "ci_low": "0.1",
      "ci_high": "0.4",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3888888888888889",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6616915422885572",
      "ci_low": "0.5904255319148937",
      "ci_high": "0.7330097087378641",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.8407960199004975",
      "ci_low": "0.7878752144082333",
      "ci_high": "0.8899533492822966",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.7869822485207101",
      "ci_low": "0.7159090909090909",
      "ci_high": "0.8529453306746426",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.1791044776119403",
      "ci_low": "0.12290502793296089",
      "ci_high": "0.24000490196078422",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.263681592039801",
      "ci_low": "0.2",
      "ci_high": "0.3297311673375503",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "75.72139303482588",
      "ci_low": "70.39761555158326",
      "ci_high": "80.9649440630562",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "49.35323383084577",
      "ci_low": "38.985007496251875",
      "ci_high": "59.52145222802087",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5266628874269781",
      "ci_low": "0.4566674409819737",
      "ci_high": "0.5557618706428814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6827111503683049",
      "ci_low": "0.59197631238404",
      "ci_high": "0.7204320545370686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "-50.0",
      "ci_low": "-100.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "10.0",
      "ci_high": "10.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "10.0",
      "ci_high": "10.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "0.0",
      "ci_high": "10.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "-45.0",
      "ci_low": "-90.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.08571428571428572",
      "ci_low": "0.02702702702702703",
      "ci_high": "0.15714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "9973",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.08571428571428572",
      "ci_low": "0.02702702702702703",
      "ci_high": "0.15714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.08571428571428572",
      "ci_low": "0.02702702702702703",
      "ci_high": "0.15714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "9.285714285714286",
      "ci_low": "5.9153329065300895",
      "ci_high": "13.225806451612902",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.7142857142857142",
      "ci_low": "-3.976506024096385",
      "ci_high": "4.821428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.044444444444444446",
      "ci_low": "0.0",
      "ci_high": "0.1111111111111111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8704",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.044444444444444446",
      "ci_low": "0.0",
      "ci_high": "0.1111111111111111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.1111111111111111",
      "ci_low": "0.020833333333333332",
      "ci_high": "0.22857142857142856",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "4.666666666666667",
      "ci_low": "2.6666666666666665",
      "ci_high": "6.984523809523804",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "-6.444444444444445",
      "ci_low": "-17.42875457875458",
      "ci_high": "2.1951219512195124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5866666666666667",
      "ci_low": "0.5112107623318386",
      "ci_high": "0.6635965365460519",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.8133333333333334",
      "ci_low": "0.755656108597285",
      "ci_high": "0.8669724770642202",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.7213114754098361",
      "ci_low": "0.6439007684597394",
      "ci_high": "0.7978142076502732",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.22666666666666666",
      "ci_low": "0.16279069767441862",
      "ci_high": "0.2916666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3377777777777778",
      "ci_low": "0.26804123711340205",
      "ci_high": "0.40969162995594716",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "71.68888888888888",
      "ci_low": "66.4621153141601",
      "ci_high": "77.05321346336177",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "37.91111111111111",
      "ci_low": "26.944377022653725",
      "ci_high": "48.924555874861205",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.47045920381833695",
      "ci_low": "0.412685902747625",
      "ci_high": "0.5001294799181797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6447033533806841",
      "ci_low": "0.5655325333948935",
      "ci_high": "0.6853626206286165",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4788732394366197",
      "ci_low": "0.3620689655172414",
      "ci_high": "0.6",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6901408450704225",
      "ci_low": "0.5789473684210527",
      "ci_high": "0.7963041431261769",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6938775510204082",
      "ci_low": "0.5555555555555556",
      "ci_high": "0.8285888501742158",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.2112676056338028",
      "ci_low": "0.11392405063291139",
      "ci_high": "0.3142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.23943661971830985",
      "ci_low": "0.13924050632911392",
      "ci_high": "0.3472222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "50.53265822784811",
      "ci_high": "69.48728813559322",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "36.056338028169016",
      "ci_low": "19.544623059866964",
      "ci_high": "52.25827956989247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.42612244897959184",
      "ci_low": "0.27095181405895696",
      "ci_high": "0.4276201814058957",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5523809523809524",
      "ci_low": "0.3512338330393886",
      "ci_high": "0.554322457378013",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.15614617940199335",
      "ci_low": "0.116363350485992",
      "ci_high": "0.19931271477663232",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.38870431893687707",
      "ci_low": "0.3354222329394743",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.4017094017094017",
      "ci_low": "0.31297709923664124",
      "ci_high": "0.491670081967213",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.23255813953488372",
      "ci_low": "0.18688439503091497",
      "ci_high": "0.28135845773936463",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2823920265780731",
      "ci_low": "0.2326388888888889",
      "ci_high": "0.33440514469453375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "27.70764119601329",
      "ci_low": "23.846091205211728",
      "ci_high": "31.74499307552999",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "-0.53156146179402",
      "ci_low": "-6.974614064329225",
      "ci_high": "5.94064150824028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.10846636868463494",
      "ci_low": "0.08329116758465686",
      "ci_high": "0.1246453781512605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16269955302695238",
      "ci_low": "0.12493675137698529",
      "ci_high": "0.18696806722689074",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.0641025641025641",
      "ci_low": "0.0136986301369863",
      "ci_high": "0.12658227848101267",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.2261808755760369",
      "ci_high": "0.4432133838383834",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.19230769230769232",
      "ci_low": "0.047619047619047616",
      "ci_high": "0.36666666666666664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.2692307692307692",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3076923076923077",
      "ci_low": "0.19753086419753085",
      "ci_high": "0.41936479490242917",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "20.76923076923077",
      "ci_low": "15.068493150684931",
      "ci_high": "27.20037974683544",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "-10.0",
      "ci_low": "-20.484119872560733",
      "ci_high": "0.6560109289617447",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.06923076923076923",
      "ci_low": "0.01282051282051282",
      "ci_high": "0.09487179487179488",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.11249999999999999",
      "ci_low": "0.020833333333333332",
      "ci_high": "0.15416666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.10493827160493827",
      "ci_low": "0.0707395498392283",
      "ci_high": "0.14285714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.24074074074074073",
      "ci_low": "0.19314534321624235",
      "ci_high": "0.29070026223173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.4358974358974359",
      "ci_low": "0.3188405797101449",
      "ci_high": "0.5569620253164557",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.13580246913580246",
      "ci_low": "0.09771940169980156",
      "ci_high": "0.1760797342192691",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.17592592592592593",
      "ci_low": "0.13354037267080746",
      "ci_high": "0.22115384615384615",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "20.15432098765432",
      "ci_low": "16.63634960718294",
      "ci_high": "23.95283851003528",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "2.5617283950617287",
      "ci_low": "-2.138373868469844",
      "ci_high": "7.43981157186887",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.12773670773670773",
      "ci_low": "0.08511994134176797",
      "ci_high": "0.1475071928071928",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.16807461544303648",
      "ci_low": "0.11199992281811574",
      "ci_high": "0.19408841158841156",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.18421052631578946",
      "ci_low": "0.09722222222222222",
      "ci_high": "0.2835820895522388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3815789473684211",
      "ci_low": "0.26582278481012656",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.4827586206896552",
      "ci_low": "0.2857142857142857",
      "ci_high": "0.68",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.19736842105263158",
      "ci_low": "0.10588235294117647",
      "ci_high": "0.2962962962962963",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.32894736842105265",
      "ci_low": "0.21516726943942135",
      "ci_high": "0.44871794871794873",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "33.94736842105263",
      "ci_low": "24.71225182863114",
      "ci_high": "43.077464788732385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "1.0526315789473684",
      "ci_low": "-11.830985915492958",
      "ci_high": "14.210896960711631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.12527472527472527",
      "ci_low": "0.0699538602423218",
      "ci_high": "0.15511197136197138",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.16285714285714287",
      "ci_low": "0.09094001831501831",
      "ci_high": "0.2016455627705628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5047619047619047",
      "ci_low": "0.4077669902912621",
      "ci_high": "0.6018540992448759",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.7333333333333333",
      "ci_low": "0.6463287601626017",
      "ci_high": "0.8151283339422725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6883116883116883",
      "ci_low": "0.5769230769230769",
      "ci_high": "0.7957044107965766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.22857142857142856",
      "ci_low": "0.14678305915849416",
      "ci_high": "0.3148148148148148",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.26666666666666666",
      "ci_low": "0.18",
      "ci_high": "0.354846246608381",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "62.95238095238095",
      "ci_low": "55.042391304347824",
      "ci_high": "70.61960730088495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "36.285714285714285",
      "ci_low": "21.362677984665936",
      "ci_high": "50.77027027027025",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4264029373404373",
      "ci_low": "0.3022962184873949",
      "ci_high": "0.4499582747350258",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5685372497872497",
      "ci_low": "0.4030616246498599",
      "ci_high": "0.5999443663133679",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4105263157894737",
      "ci_low": "0.35018050541516244",
      "ci_high": "0.4725274725274725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.6982456140350877",
      "ci_low": "0.6420620574448987",
      "ci_high": "0.7526501766784452",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5879396984924623",
      "ci_low": "0.5149962871287129",
      "ci_high": "0.6614583333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.28771929824561404",
      "ci_low": "0.23371548761701624",
      "ci_high": "0.3448303142732431",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.36140350877192984",
      "ci_low": "0.3028161999887773",
      "ci_high": "0.4204946996466431",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "54.35087719298245",
      "ci_low": "49.52229027629834",
      "ci_high": "59.32211797148238",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "18.210526315789476",
      "ci_low": "9.029823902072373",
      "ci_high": "27.44281077617429",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.31125373401485507",
      "ci_low": "0.26469666469774217",
      "ci_high": "0.32501670601740834",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.45490930356017284",
      "ci_low": "0.3868643560967001",
      "ci_high": "0.47502441648698135",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.24897959183673468",
      "ci_low": "0.19343028145745098",
      "ci_high": "0.3080168776371308",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.40816326530612246",
      "ci_low": "0.34321723636621515",
      "ci_high": "0.47388097164778514",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.61",
      "ci_low": "0.5094295946890287",
      "ci_high": "0.7096774193548387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.15918367346938775",
      "ci_low": "0.11290165683444207",
      "ci_high": "0.20779348102877515",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.1836734693877551",
      "ci_low": "0.13461322869955156",
      "ci_high": "0.23577235772357724",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "36.0",
      "ci_low": "30.753916040100254",
      "ci_high": "41.411775931746746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "17.63265306122449",
      "ci_low": "10.730401056279915",
      "ci_high": "24.85131501182033",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.27746545691751173",
      "ci_low": "0.19466994371321317",
      "ci_high": "0.2806427039474985",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.3433047178809891",
      "ci_low": "0.24086281171295867",
      "ci_high": "0.34723588793504045",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.46923076923076923",
      "ci_low": "0.38297456781914896",
      "ci_high": "0.5581395348837209",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7076923076923077",
      "ci_low": "0.6260851891127608",
      "ci_high": "0.7867647058823529",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6630434782608695",
      "ci_low": "0.5652173913043478",
      "ci_high": "0.7592681623931623",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.23846153846153847",
      "ci_low": "0.16778523489932887",
      "ci_high": "0.31297709923664124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.3153846153846154",
      "ci_low": "0.23577235772357724",
      "ci_high": "0.3969465648854962",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "60.38461538461539",
      "ci_low": "53.3883646775558",
      "ci_high": "67.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "28.846153846153847",
      "ci_low": "16.837398373983742",
      "ci_high": "41.19062881562881",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.39531404268246373",
      "ci_low": "0.2764176065162907",
      "ci_high": "0.3891403382192857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5121113734750098",
      "ci_low": "0.35808644480519486",
      "ci_high": "0.5041136199658927",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.11842105263157894",
      "ci_low": "0.05063291139240506",
      "ci_high": "0.19480519480519481",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.2631578947368421",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.3625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.45",
      "ci_low": "0.22727272727272727",
      "ci_high": "0.6666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.14473684210526316",
      "ci_low": "0.07017543859649122",
      "ci_high": "0.22727272727272727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14544721407624633",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "19.605263157894736",
      "ci_low": "12.592592592592593",
      "ci_high": "27.143339768339764",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "-4.078947368421052",
      "ci_low": "-15.915631949592292",
      "ci_high": "7.826086956521739",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.16049382716049382",
      "ci_low": "0.05555555555555555",
      "ci_high": "0.14814814814814814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.18840579710144925",
      "ci_low": "0.06521739130434782",
      "ci_high": "0.17391304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.30612244897959184",
      "ci_low": "0.21176470588235294",
      "ci_high": "0.40426129572077446",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4897959183673469",
      "ci_low": "0.38823529411764707",
      "ci_high": "0.59",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.625",
      "ci_low": "0.4782608695652174",
      "ci_high": "0.761924939467312",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.1836734693877551",
      "ci_low": "0.11008944954128441",
      "ci_high": "0.2631578947368421",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2653061224489796",
      "ci_low": "0.18018018018018017",
      "ci_high": "0.35714285714285715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "41.53061224489796",
      "ci_low": "33.13095238095239",
      "ci_high": "50.180626170725176",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "15.0",
      "ci_low": "1.276543596161869",
      "ci_high": "28.449712299208823",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.24488536155202822",
      "ci_low": "0.1366547958214625",
      "ci_high": "0.24438029563029562",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2874741200828157",
      "ci_low": "0.16042084726867334",
      "ci_high": "0.28688121660947746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.1588785046728972",
      "ci_low": "0.09174120795107034",
      "ci_high": "0.23232323232323232",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.32710280373831774",
      "ci_low": "0.2389279045222051",
      "ci_high": "0.41818181818181815",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.4857142857142857",
      "ci_low": "0.3142857142857143",
      "ci_high": "0.6538461538461539",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.16822429906542055",
      "ci_low": "0.10084033613445378",
      "ci_high": "0.2413793103448276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.205607476635514",
      "ci_low": "0.13157894736842105",
      "ci_high": "0.2868870277975766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "26.35514018691589",
      "ci_low": "19.652173913043477",
      "ci_high": "33.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "5.794392523364485",
      "ci_low": "-4.00042735042735",
      "ci_high": "15.666752577319587",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.15526932084309134",
      "ci_low": "0.07467896174863388",
      "ci_high": "0.16159250585480095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.18571428571428572",
      "ci_low": "0.0893218954248366",
      "ci_high": "0.19327731092436973",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3007518796992481",
      "ci_low": "0.2222222222222222",
      "ci_high": "0.3821138211382114",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6616541353383458",
      "ci_low": "0.5781163194444444",
      "ci_high": "0.7416666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.45454545454545453",
      "ci_low": "0.34615384615384615",
      "ci_high": "0.5643564356435643",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3609022556390977",
      "ci_low": "0.2773109243697479",
      "ci_high": "0.44696969696969696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.45864661654135336",
      "ci_low": "0.3700787401574803",
      "ci_high": "0.5447761194029851",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "46.2406015037594",
      "ci_low": "39.66081096986817",
      "ci_high": "53.00018796992481",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.37593984962406013",
      "ci_low": "-11.971136021318504",
      "ci_high": "13.35992299910071",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.21611721611721613",
      "ci_low": "0.14600068232421173",
      "ci_high": "0.22378513831568128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.28668610301263364",
      "ci_low": "0.19367437451170944",
      "ci_high": "0.2968578365412099",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / primary_blind",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4592074592074592",
      "ci_low": "0.4110552226027397",
      "ci_high": "0.5072824413923751",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4592074592074592",
      "ci_low": "0.4110552226027397",
      "ci_high": "0.5072824413923751",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5407925407925408",
      "ci_low": "0.4927175586076249",
      "ci_high": "0.5889447773972603",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.298983907576611",
      "ci_low": "0.2477162346035455",
      "ci_high": "0.30748940186481716",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.49669907226436993",
      "ci_low": "0.4115285832929869",
      "ci_high": "0.5108291676141319",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4155124653739612",
      "ci_low": "0.3628755309325946",
      "ci_high": "0.46978356528857956",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4155124653739612",
      "ci_low": "0.3628755309325946",
      "ci_high": "0.46978356528857956",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5844875346260388",
      "ci_low": "0.5302164347114204",
      "ci_high": "0.6371244690674053",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.27176729202217",
      "ci_low": "0.20811538906329702",
      "ci_high": "0.2800955122292008",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.39613537481197664",
      "ci_low": "0.3033546349058227",
      "ci_high": "0.4082748144357842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6026666666666667",
      "ci_low": "0.5524296675191815",
      "ci_high": "0.6520547945205479",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6026666666666667",
      "ci_low": "0.5524296675191815",
      "ci_high": "0.6520547945205479",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3973333333333333",
      "ci_low": "0.34794520547945207",
      "ci_high": "0.4475703324808184",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.44580929226526594",
      "ci_low": "0.38067166726259244",
      "ci_high": "0.4522095006908864",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6001278934340117",
      "ci_low": "0.512442629007336",
      "ci_high": "0.6087435586223469",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1724137931034483",
      "ci_low": "0.11666666666666667",
      "ci_high": "0.23115645431458368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1724137931034483",
      "ci_low": "0.11666666666666667",
      "ci_high": "0.23115645431458368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.8275862068965517",
      "ci_low": "0.7688435456854164",
      "ci_high": "0.8833333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.10125866241293545",
      "ci_low": "0.05421660362698098",
      "ci_high": "0.10816291556553269",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1393950417632618",
      "ci_low": "0.07463584395402577",
      "ci_high": "0.1488995980512528",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.375",
      "ci_low": "0.31336257194280176",
      "ci_high": "0.4393305439330544",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.375",
      "ci_low": "0.31336257194280176",
      "ci_high": "0.4393305439330544",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.625",
      "ci_low": "0.5606694560669456",
      "ci_high": "0.6866374280571982",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2610571617924559",
      "ci_low": "0.19009548329839487",
      "ci_high": "0.2654312600239458",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3370611456054494",
      "ci_low": "0.24543973792957313",
      "ci_high": "0.34270871547395537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4175",
      "ci_low": "0.3686746987951807",
      "ci_high": "0.46594142718893056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4175",
      "ci_low": "0.3686746987951807",
      "ci_high": "0.46594142718893056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5825",
      "ci_low": "0.5340585728110694",
      "ci_high": "0.6313253012048192",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2078819215356367",
      "ci_low": "0.18175755511458577",
      "ci_high": "0.2205133767698909",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5197048038390918",
      "ci_low": "0.4543938877864644",
      "ci_high": "0.5512834419247273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.46153846153846156",
      "ci_low": "0.4083742576132689",
      "ci_high": "0.5144375999833795",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.46153846153846156",
      "ci_low": "0.4083742576132689",
      "ci_high": "0.5144375999833795",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5384615384615384",
      "ci_low": "0.4855624000166206",
      "ci_high": "0.591625742386731",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.2874866981663751",
      "ci_low": "0.2507405535490119",
      "ci_high": "0.3077505863215063",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4644015893456829",
      "ci_low": "0.40504243265609624",
      "ci_high": "0.4971355625193564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.43853820598006643",
      "ci_low": "0.38235294117647056",
      "ci_high": "0.4952681388012618",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.43853820598006643",
      "ci_low": "0.38235294117647056",
      "ci_high": "0.4952681388012618",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5614617940199336",
      "ci_low": "0.5047318611987381",
      "ci_high": "0.6176470588235294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.3220388096190407",
      "ci_low": "0.2632268989412078",
      "ci_high": "0.3407247790468266",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.458905303707133",
      "ci_low": "0.37509833099122114",
      "ci_high": "0.48553281014172783",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3925925925925926",
      "ci_low": "0.32986019736842104",
      "ci_high": "0.4574468085106383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3925925925925926",
      "ci_low": "0.32986019736842104",
      "ci_high": "0.4574468085106383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6074074074074074",
      "ci_low": "0.5425531914893617",
      "ci_high": "0.6701398026315789",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.21514818860131124",
      "ci_low": "0.17979880126996894",
      "ci_high": "0.23647038880357946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.39842257148390975",
      "ci_low": "0.3329607430925351",
      "ci_high": "0.437908127414036",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4977168949771689",
      "ci_low": "0.42924528301886794",
      "ci_high": "0.5662100456621004",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4977168949771689",
      "ci_low": "0.42924528301886794",
      "ci_high": "0.5662100456621004",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.502283105022831",
      "ci_low": "0.4337899543378995",
      "ci_high": "0.5707547169811321",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.29333447516465333",
      "ci_low": "0.2370585608206545",
      "ci_high": "0.3121717862304542",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5422243328801167",
      "ci_low": "0.43819915788060376",
      "ci_high": "0.5770448169714456",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.4315068493150685",
      "ci_low": "0.375",
      "ci_high": "0.48905109489051096",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.4315068493150685",
      "ci_low": "0.375",
      "ci_high": "0.48905109489051096",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.5684931506849316",
      "ci_low": "0.5109489051094891",
      "ci_high": "0.625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.32647693293255553",
      "ci_low": "0.2596057519303023",
      "ci_high": "0.3428253231345401",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.46184541731922496",
      "ci_low": "0.36724716126725687",
      "ci_high": "0.4849724083366665",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5263157894736842",
      "ci_low": "0.45077720207253885",
      "ci_high": "0.6011560693641619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5263157894736842",
      "ci_low": "0.45077720207253885",
      "ci_high": "0.6011560693641619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.47368421052631576",
      "ci_low": "0.3988439306358382",
      "ci_high": "0.5492227979274611",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.2788979274693561",
      "ci_low": "0.21340292085361107",
      "ci_high": "0.28763592454464304",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.47616719324036405",
      "ci_low": "0.36434645023787254",
      "ci_high": "0.4910857248323173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.30337078651685395",
      "ci_low": "0.20238095238095238",
      "ci_high": "0.4084507042253521",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.30337078651685395",
      "ci_low": "0.20238095238095238",
      "ci_high": "0.4084507042253521",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6966292134831461",
      "ci_low": "0.5915492957746479",
      "ci_high": "0.7976190476190477",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.1948786507610037",
      "ci_low": "0.09221547350997815",
      "ci_high": "0.20003545090050276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.4517641449459631",
      "ci_low": "0.21377223404585843",
      "ci_high": "0.4637185452693473",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.32142857142857145",
      "ci_low": "0.07142857142857142",
      "ci_high": "0.35714285714285715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.4228855721393035",
      "ci_low": "0.35428507367072387",
      "ci_high": "0.49484667578371555",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.4228855721393035",
      "ci_low": "0.35428507367072387",
      "ci_high": "0.49484667578371555",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5771144278606966",
      "ci_low": "0.5051533242162844",
      "ci_high": "0.6457149263292761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.2632130605408002",
      "ci_low": "0.19944698733644303",
      "ci_high": "0.2855327072917179",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.43868843423466697",
      "ci_low": "0.3324116455607384",
      "ci_high": "0.4758878454861966",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.2",
      "ci_high": "0.43103448275862066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.2",
      "ci_high": "0.43103448275862066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6857142857142857",
      "ci_low": "0.5689655172413793",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.10008018546917977",
      "ci_low": "0.06625998647645506",
      "ci_high": "0.12302532508604615",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.44321224993493896",
      "ci_low": "0.29343708296715815",
      "ci_high": "0.5448264396667758",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.2826086956521739",
      "ci_high": "0.5714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.2826086956521739",
      "ci_high": "0.5714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5777777777777777",
      "ci_low": "0.42857142857142855",
      "ci_high": "0.717391304347826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.25553952220618886",
      "ci_low": "0.1253138528138528",
      "ci_high": "0.2731054829339143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5307359307359308",
      "ci_low": "0.2602672327672328",
      "ci_high": "0.5672190799396681",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.37333333333333335",
      "ci_low": "0.305415994594359",
      "ci_high": "0.4458874458874459",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.37333333333333335",
      "ci_low": "0.305415994594359",
      "ci_high": "0.4458874458874459",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6266666666666667",
      "ci_low": "0.5541125541125541",
      "ci_high": "0.694584005405641",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.21979367022637283",
      "ci_low": "0.17855857692814214",
      "ci_high": "0.24327717807615812",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.39074430262466275",
      "ci_low": "0.31743747009447487",
      "ci_high": "0.43249276102428125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.49295774647887325",
      "ci_high": "0.7162162162162162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.49295774647887325",
      "ci_high": "0.7162162162162162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.39436619718309857",
      "ci_low": "0.28378378378378377",
      "ci_high": "0.5070422535211268",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.45614368796186977",
      "ci_low": "0.29595723720723716",
      "ci_high": "0.47083858061130784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5575089519533963",
      "ci_low": "0.3617255121421788",
      "ci_high": "0.5754693763027097",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.43853820598006643",
      "ci_low": "0.38235294117647056",
      "ci_high": "0.4952681388012618",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.43853820598006643",
      "ci_low": "0.38235294117647056",
      "ci_high": "0.4952681388012618",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5614617940199336",
      "ci_low": "0.5047318611987381",
      "ci_high": "0.6176470588235294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.3220388096190407",
      "ci_low": "0.2632268989412078",
      "ci_high": "0.3407247790468266",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.458905303707133",
      "ci_low": "0.37509833099122114",
      "ci_high": "0.48553281014172783",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6025641025641025",
      "ci_low": "0.48717948717948717",
      "ci_high": "0.7142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6025641025641025",
      "ci_low": "0.48717948717948717",
      "ci_high": "0.7142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3974358974358974",
      "ci_low": "0.2857142857142857",
      "ci_high": "0.5128205128205128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3632527873129377",
      "ci_low": "0.26603432471473437",
      "ci_high": "0.39683367665247915",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.635692377797641",
      "ci_low": "0.46556006825078516",
      "ci_high": "0.6944589341418386",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.4444444444444444",
      "ci_low": "0.39062127976190475",
      "ci_high": "0.49847140996147343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.4444444444444444",
      "ci_low": "0.39062127976190475",
      "ci_high": "0.49847140996147343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5555555555555556",
      "ci_low": "0.5015285900385266",
      "ci_high": "0.6093787202380951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.2387020889204657",
      "ci_low": "0.2043271972044104",
      "ci_high": "0.2530410745582378",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5276572491926084",
      "ci_low": "0.4516706464518546",
      "ci_high": "0.559353954286631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3026315789473684",
      "ci_low": "0.19480519480519481",
      "ci_high": "0.4142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3026315789473684",
      "ci_low": "0.19480519480519481",
      "ci_high": "0.4142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6973684210526315",
      "ci_low": "0.5857142857142857",
      "ci_high": "0.8051948051948052",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.16010989756790428",
      "ci_low": "0.07980530090824207",
      "ci_high": "0.17705178686561662",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3762582592845751",
      "ci_low": "0.18754245713436893",
      "ci_high": "0.41607169913419906",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5047619047619047",
      "ci_low": "0.4019607843137255",
      "ci_high": "0.6068376068376068",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5047619047619047",
      "ci_low": "0.4019607843137255",
      "ci_high": "0.6068376068376068",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.49523809523809526",
      "ci_low": "0.39316239316239315",
      "ci_high": "0.5980392156862745",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.3514661528947243",
      "ci_low": "0.22981918280132563",
      "ci_high": "0.37307821741136643",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5125548063048063",
      "ci_low": "0.33515297491859997",
      "ci_high": "0.544072400391576",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4456140350877193",
      "ci_low": "0.38571428571428573",
      "ci_high": "0.5076335877862596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4456140350877193",
      "ci_low": "0.38571428571428573",
      "ci_high": "0.5076335877862596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5543859649122806",
      "ci_low": "0.49236641221374045",
      "ci_high": "0.6142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3042640295820698",
      "ci_low": "0.2486817262463597",
      "ci_high": "0.32480684543414234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4212886563444043",
      "ci_low": "0.344328544033421",
      "ci_high": "0.4497325552165048",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6204081632653061",
      "ci_low": "0.5575202410092261",
      "ci_high": "0.6816326530612244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6204081632653061",
      "ci_low": "0.5575202410092261",
      "ci_high": "0.6816326530612244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.3795918367346939",
      "ci_low": "0.3183673469387755",
      "ci_high": "0.44247975899077385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.4624122364563541",
      "ci_low": "0.3679518601190476",
      "ci_high": "0.46041481782727794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6269996426526834",
      "ci_low": "0.49891777643260693",
      "ci_high": "0.6242912784098683",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5692307692307692",
      "ci_low": "0.4827586206896552",
      "ci_high": "0.656934306569343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5692307692307692",
      "ci_low": "0.4827586206896552",
      "ci_high": "0.656934306569343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4307692307692308",
      "ci_low": "0.34306569343065696",
      "ci_high": "0.5172413793103449",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4355060613989185",
      "ci_low": "0.31235756802721093",
      "ci_high": "0.4413523158389229",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5542804417804418",
      "ci_low": "0.39754599567099574",
      "ci_high": "0.5617211292495383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.09210526315789473",
      "ci_low": "0.03125",
      "ci_high": "0.16216216216216217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.09210526315789473",
      "ci_low": "0.03125",
      "ci_high": "0.16216216216216217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.9078947368421053",
      "ci_low": "0.8378378378378378",
      "ci_high": "0.96875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.05592020592020591",
      "ci_low": "0.014414414414414415",
      "ci_high": "0.06516545391545392",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.08995859213250516",
      "ci_low": "0.02318840579710145",
      "ci_high": "0.10483138238573021",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.23469387755102042",
      "ci_low": "0.15217391304347827",
      "ci_high": "0.32222222222222224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.23469387755102042",
      "ci_low": "0.15217391304347827",
      "ci_high": "0.32222222222222224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.7653061224489796",
      "ci_low": "0.6777777777777778",
      "ci_high": "0.8478260869565217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.14702956424267902",
      "ci_low": "0.07613816738816738",
      "ci_high": "0.16141058167697508",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.19497398736529173",
      "ci_low": "0.10096583066691761",
      "ci_high": "0.2140444670064235",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.34579439252336447",
      "ci_low": "0.25742574257425743",
      "ci_high": "0.4347909060736806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.34579439252336447",
      "ci_low": "0.25742574257425743",
      "ci_high": "0.4347909060736806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6542056074766355",
      "ci_low": "0.5652090939263192",
      "ci_high": "0.7425742574257426",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2116708908375575",
      "ci_low": "0.13088819745069744",
      "ci_high": "0.2191927208593875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.29882949294714",
      "ci_low": "0.18478333757745521",
      "ci_high": "0.30944854709560593",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.39849624060150374",
      "ci_low": "0.3115912330719886",
      "ci_high": "0.4838709677419355",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.39849624060150374",
      "ci_low": "0.3115912330719886",
      "ci_high": "0.4838709677419355",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6015037593984962",
      "ci_low": "0.5161290322580645",
      "ci_high": "0.6884087669280113",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.29384673706707604",
      "ci_low": "0.19188352960598723",
      "ci_high": "0.29939263662522686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.35381545891749977",
      "ci_low": "0.23104343360720914",
      "ci_high": "0.3604931747120078",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "BioCLIP 2.5 / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.21212121212121213",
      "ci_low": "0.17084282460136674",
      "ci_high": "0.2548555587283457",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.28438228438228436",
      "ci_low": "0.2399976247030879",
      "ci_high": "0.33049834700954095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7459016393442623",
      "ci_low": "0.6607142857142857",
      "ci_high": "0.8260869565217391",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.07226107226107226",
      "ci_low": "0.04741379310344827",
      "ci_high": "0.09887640449438202",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.13286713286713286",
      "ci_low": "0.10048962341946124",
      "ci_high": "0.16625916870415647",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "30.792540792540795",
      "ci_low": "26.837150629400472",
      "ci_high": "34.890613514760936",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "17.505827505827504",
      "ci_low": "12.414399308192019",
      "ci_high": "22.69328676026828",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.3155633928361201",
      "ci_low": "0.24246509857023862",
      "ci_high": "0.32642063674260646",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.3359223214061924",
      "ci_low": "0.25810800815541524",
      "ci_high": "0.34748003266148425",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3684210526315789",
      "ci_low": "0.31666666666666665",
      "ci_high": "0.4231807883120434",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.42382271468144045",
      "ci_low": "0.3693170332026464",
      "ci_high": "0.47863475177304965",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.869281045751634",
      "ci_low": "0.8092105263157895",
      "ci_high": "0.9246575342465754",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.055401662049861494",
      "ci_low": "0.03154506340784395",
      "ci_high": "0.08256948391112554",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.19390581717451524",
      "ci_low": "0.15193282747045248",
      "ci_high": "0.23782234957020057",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "50.02770083102493",
      "ci_low": "45.58192619283124",
      "ci_high": "54.49860315186247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "30.63711911357341",
      "ci_low": "23.95265058175663",
      "ci_high": "37.3644866333725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.34954589419670884",
      "ci_low": "0.29738036619515024",
      "ci_high": "0.36543287850737854",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.36139490755930914",
      "ci_low": "0.30746105657464684",
      "ci_high": "0.3778204337110185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.33866666666666667",
      "ci_low": "0.28895026754800124",
      "ci_high": "0.39071126048295335",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.4106666666666667",
      "ci_low": "0.35883700043230027",
      "ci_high": "0.46329422661315217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8246753246753247",
      "ci_low": "0.7579580513535032",
      "ci_high": "0.8875828912466842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.072",
      "ci_low": "0.04497354497354497",
      "ci_high": "0.10084033613445378",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.13333333333333333",
      "ci_low": "0.09755784865540963",
      "ci_high": "0.17187594582324453",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "44.69333333333333",
      "ci_low": "40.15383631713555",
      "ci_high": "49.24534871498397",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "31.36",
      "ci_low": "25.43303089651865",
      "ci_high": "37.358538454171054",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3580125159393452",
      "ci_low": "0.2972066855908319",
      "ci_high": "0.36557131269329507",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3763721321413629",
      "ci_low": "0.3124480540826695",
      "ci_high": "0.3843185594980794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.22413793103448276",
      "ci_low": "0.15950920245398773",
      "ci_high": "0.29081864564007415",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.25287356321839083",
      "ci_low": "0.18681318681318682",
      "ci_high": "0.32085891595695515",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.8863636363636364",
      "ci_low": "0.78",
      "ci_high": "0.9736842105263158",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.028735632183908046",
      "ci_low": "0.005952380952380952",
      "ci_high": "0.056179775280898875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.15517241379310345",
      "ci_low": "0.10344827586206896",
      "ci_high": "0.21142857142857144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "31.724137931034484",
      "ci_low": "25.67540228089361",
      "ci_high": "37.86526063731811",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "16.206896551724135",
      "ci_low": "7.575701871657754",
      "ci_high": "24.761904761904763",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2171589310829817",
      "ci_low": "0.129957805907173",
      "ci_high": "0.21377347160575008",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2227994227994228",
      "ci_low": "0.13333333333333333",
      "ci_high": "0.21932602930979556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.24166666666666667",
      "ci_low": "0.18749377490039842",
      "ci_high": "0.298702645901816",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.32083333333333336",
      "ci_low": "0.25957446808510637",
      "ci_high": "0.38376795598313124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.7532467532467533",
      "ci_low": "0.6385542168674698",
      "ci_high": "0.8589743589743589",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.07916666666666666",
      "ci_low": "0.042550682456844645",
      "ci_high": "0.12174187066130797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.20833333333333334",
      "ci_low": "0.15596155858652808",
      "ci_high": "0.26337448559670784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "37.416666666666664",
      "ci_low": "32.499748995983936",
      "ci_high": "42.54191100823045",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "16.583333333333336",
      "ci_low": "9.132100223009314",
      "ci_high": "24.138039243976337",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.23777415999638224",
      "ci_low": "0.17657195428028763",
      "ci_high": "0.24916219068996848",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.24379375898363242",
      "ci_low": "0.1810421303380164",
      "ci_high": "0.25547009425173983",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1225",
      "ci_low": "0.08844842347228982",
      "ci_high": "0.15931545818841672",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2025",
      "ci_low": "0.16058394160583941",
      "ci_high": "0.24611498605021917",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6049382716049383",
      "ci_low": "0.4861111111111111",
      "ci_high": "0.7222222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.08",
      "ci_low": "0.05223880597014925",
      "ci_high": "0.11027659075746919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1575",
      "ci_low": "0.11893171083705065",
      "ci_high": "0.19845360824742267",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "22.05",
      "ci_low": "18.620645559573155",
      "ci_high": "25.60983917059879",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "6.3",
      "ci_low": "1.8015145616731663",
      "ci_high": "10.899755440051537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.13656107706909845",
      "ci_low": "0.10622364253569397",
      "ci_high": "0.1583469879089808",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.15021718477600832",
      "ci_low": "0.11684600678926339",
      "ci_high": "0.17418168669987888",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.44871794871794873",
      "ci_low": "0.3969839656296751",
      "ci_high": "0.502591010902418",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.49743589743589745",
      "ci_low": "0.44472953216374267",
      "ci_high": "0.551076441959617",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.9020618556701031",
      "ci_low": "0.8502355072463769",
      "ci_high": "0.9476792406039898",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.04871794871794872",
      "ci_low": "0.02564102564102564",
      "ci_high": "0.07653061224489796",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.1641025641025641",
      "ci_low": "0.12698412698412698",
      "ci_high": "0.20308483290488433",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "57.56410256410256",
      "ci_low": "53.350515463917525",
      "ci_high": "61.82284018987342",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "41.153846153846146",
      "ci_low": "34.55799306514733",
      "ci_high": "47.73444444444443",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4789549804666704",
      "ci_low": "0.43602614790514854",
      "ci_high": "0.5059200301212662",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4912358774017132",
      "ci_low": "0.4472063055437421",
      "ci_high": "0.518892338585914",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16611295681063123",
      "ci_low": "0.125",
      "ci_high": "0.21017355682031388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.21262458471760798",
      "ci_low": "0.167192429022082",
      "ci_high": "0.2607260726072607",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.78125",
      "ci_low": "0.6727272727272727",
      "ci_high": "0.8787878787878788",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.046511627906976744",
      "ci_low": "0.024052923315915836",
      "ci_high": "0.07165134402873755",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16279069767441862",
      "ci_low": "0.12244897959183673",
      "ci_high": "0.20569620253164558",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "27.774086378737543",
      "ci_low": "23.926013961388684",
      "ci_high": "31.83334775086505",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "11.49501661129568",
      "ci_low": "5.460672759877365",
      "ci_high": "17.45966998980028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16558021589202485",
      "ci_low": "0.12916420356021197",
      "ci_high": "0.1882910338776659",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.17385922668662604",
      "ci_low": "0.13562241373822254",
      "ci_high": "0.19770558557154919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6111111111111112",
      "ci_low": "0.545801177194307",
      "ci_high": "0.6791686320754717",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.7296296296296296",
      "ci_low": "0.6707304829072388",
      "ci_high": "0.7890909090909091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.8375634517766497",
      "ci_low": "0.7747233400402415",
      "ci_high": "0.896551724137931",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.11851851851851852",
      "ci_low": "0.07473044032506507",
      "ci_high": "0.1655183831925097",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.22592592592592592",
      "ci_low": "0.17120363657883447",
      "ci_high": "0.28188349681638264",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "72.77777777777777",
      "ci_low": "68.09911958801037",
      "ci_high": "77.59658168913693",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "50.18518518518518",
      "ci_low": "41.52919928427645",
      "ci_high": "58.839131374243735",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6101630065579071",
      "ci_low": "0.5543469641645767",
      "ci_high": "0.6483455568446482",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6553602663029373",
      "ci_low": "0.5954097022508416",
      "ci_high": "0.6963711536479554",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.0410958904109589",
      "ci_low": "0.017241379310344827",
      "ci_high": "0.0700023364485981",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.0639269406392694",
      "ci_low": "0.03317241784497278",
      "ci_high": "0.09895962252475245",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6428571428571429",
      "ci_low": "0.36363636363636365",
      "ci_high": "0.9",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.0228310502283105",
      "ci_low": "0.004651162790697674",
      "ci_high": "0.04484304932735426",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.0776255707762557",
      "ci_low": "0.042252035419266656",
      "ci_high": "0.11764705882352941",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "14.885844748858448",
      "ci_low": "11.851851851851851",
      "ci_high": "18.093023255813954",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "7.123287671232877",
      "ci_low": "2.4509037990196076",
      "ci_high": "11.626873300592289",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1019047619047619",
      "ci_low": "0.03673469387755102",
      "ci_high": "0.125578231292517",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.10808080808080807",
      "ci_low": "0.03896103896103896",
      "ci_high": "0.1331890331890332",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.17123287671232876",
      "ci_low": "0.12903225806451613",
      "ci_high": "0.21631205673758866",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.21575342465753425",
      "ci_low": "0.16974169741697417",
      "ci_high": "0.26440677966101694",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7936507936507936",
      "ci_low": "0.6875",
      "ci_high": "0.890625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.04452054794520548",
      "ci_low": "0.022058823529411766",
      "ci_high": "0.06944444444444445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.16095890410958905",
      "ci_low": "0.120271574806018",
      "ci_high": "0.2043378272782124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "28.35616438356164",
      "ci_low": "24.376891784741613",
      "ci_high": "32.516790094885444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "12.26027397260274",
      "ci_low": "6.132138948449923",
      "ci_high": "18.338983050847457",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.1854614833411003",
      "ci_low": "0.1332864670302521",
      "ci_high": "0.20475336259421725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.19450838496749545",
      "ci_low": "0.13978824590977665",
      "ci_high": "0.21474133150125224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.11052631578947368",
      "ci_low": "0.06451612903225806",
      "ci_high": "0.16326530612244897",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.13157894736842105",
      "ci_low": "0.08162988021295474",
      "ci_high": "0.18686868686868688",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.84",
      "ci_low": "0.6666666666666666",
      "ci_high": "0.9666666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.021052631578947368",
      "ci_low": "0.004901960784313725",
      "ci_high": "0.043903779147681565",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.1",
      "ci_low": "0.05747126436781609",
      "ci_high": "0.14705882352941177",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "22.526315789473685",
      "ci_low": "17.853658536585368",
      "ci_high": "27.62392661067054",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "12.526315789473683",
      "ci_low": "5.6936767860089095",
      "ci_high": "19.37853107344633",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.19259259259259262",
      "ci_low": "0.09844733044733044",
      "ci_high": "0.2025648795648796",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.2113821138211382",
      "ci_low": "0.10805194805194807",
      "ci_high": "0.22232730683950197",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.14606741573033707",
      "ci_low": "0.07142857142857142",
      "ci_high": "0.22988505747126436",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3146067415730337",
      "ci_low": "0.20833333333333334",
      "ci_high": "0.4230769230769231",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.4642857142857143",
      "ci_low": "0.26666666666666666",
      "ci_high": "0.6666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.16853932584269662",
      "ci_low": "0.08641975308641975",
      "ci_high": "0.2597402597402597",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.23595505617977527",
      "ci_low": "0.14457831325301204",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "29.887640449438205",
      "ci_low": "22.209264572636663",
      "ci_high": "37.90003086419753",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "6.2921348314606735",
      "ci_low": "-4.0846670934699105",
      "ci_high": "17.142857142857142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3294871794871795",
      "ci_low": "0.12179487179487179",
      "ci_high": "0.33504807692307687",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3893939393939394",
      "ci_low": "0.14393939393939392",
      "ci_high": "0.39596590909090895",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.125",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.125",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "9926",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "51.66666666666667",
      "ci_low": "25.714285714285715",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "43.33333333333333",
      "ci_low": "13.333333333333332",
      "ci_high": "75.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3888888888888889",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3888888888888889",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6716417910447762",
      "ci_low": "0.6011552575176983",
      "ci_high": "0.743455497382199",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.7611940298507462",
      "ci_low": "0.6984028760018859",
      "ci_high": "0.8226617778772951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.8823529411764706",
      "ci_low": "0.8223684210526315",
      "ci_high": "0.9363135821297209",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.08955223880597014",
      "ci_low": "0.048456373454597135",
      "ci_high": "0.135678391959799",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.18407960199004975",
      "ci_low": "0.12857142857142856",
      "ci_high": "0.24083948119942883",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "76.61691542288557",
      "ci_low": "71.3823650427913",
      "ci_high": "81.8100426699177",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "58.2089552238806",
      "ci_low": "49.238488530137865",
      "ci_high": "67.1041957067187",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.656634663096177",
      "ci_low": "0.5585117487786446",
      "ci_high": "0.6842099790236114",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.7052742677699678",
      "ci_low": "0.5998829894289146",
      "ci_high": "0.7348921996920271",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "-50.0",
      "ci_low": "-100.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "10.0",
      "ci_high": "10.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "10.0",
      "ci_high": "10.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "0.0",
      "ci_high": "10.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "-45.0",
      "ci_low": "-90.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.05714285714285714",
      "ci_low": "0.012341915085817527",
      "ci_high": "0.11940298507462686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "9801",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.05714285714285714",
      "ci_low": "0.012341915085817527",
      "ci_high": "0.11940298507462686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.07142857142857142",
      "ci_low": "0.014925373134328358",
      "ci_high": "0.13924050632911392",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "9.142857142857142",
      "ci_low": "5.810701830863121",
      "ci_high": "13.064516129032258",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "2.0",
      "ci_low": "-2.3189052795031055",
      "ci_high": "5.647186700767261",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.022222222222222223",
      "ci_low": "0.0",
      "ci_high": "0.07142857142857142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "6322",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.022222222222222223",
      "ci_low": "0.0",
      "ci_high": "0.07142857142857142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.1111111111111111",
      "ci_low": "0.020833333333333332",
      "ci_high": "0.22857142857142856",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "4.666666666666667",
      "ci_low": "2.6666666666666665",
      "ci_high": "6.984523809523804",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "-6.444444444444445",
      "ci_low": "-17.42875457875458",
      "ci_high": "2.1951219512195124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.5254131355932203",
      "ci_high": "0.6755325059101654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.7377777777777778",
      "ci_low": "0.6729835673473004",
      "ci_high": "0.8026330540476753",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.8132530120481928",
      "ci_low": "0.7401101694915254",
      "ci_high": "0.8807986439608955",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.13777777777777778",
      "ci_low": "0.08620689655172414",
      "ci_high": "0.1924916301085199",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.2577777777777778",
      "ci_low": "0.19457013574660634",
      "ci_high": "0.32367149758454106",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "72.17777777777778",
      "ci_low": "66.88932585244025",
      "ci_high": "77.60014084507041",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "46.4",
      "ci_low": "36.288182584634015",
      "ci_high": "56.138828168531134",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.607383092987342",
      "ci_low": "0.5358051339869698",
      "ci_high": "0.6445723300162344",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6523744332086266",
      "ci_low": "0.5754944031711898",
      "ci_high": "0.6923184285359554",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4788732394366197",
      "ci_low": "0.3620689655172414",
      "ci_high": "0.6",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.49295774647887325",
      "ci_low": "0.375",
      "ci_high": "0.6129032258064516",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.9714285714285714",
      "ci_low": "0.9032258064516129",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.014084507042253521",
      "ci_low": "0.0",
      "ci_high": "0.046875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.07042253521126761",
      "ci_low": "0.015151515151515152",
      "ci_high": "0.13559845155890346",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "60.42253521126761",
      "ci_low": "51.12676056338028",
      "ci_high": "69.8360655737705",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "53.38028169014085",
      "ci_low": "40.62475961538462",
      "ci_high": "65.40568693693693",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5523809523809524",
      "ci_low": "0.3512338330393886",
      "ci_high": "0.554322457378013",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5523809523809524",
      "ci_low": "0.3512338330393886",
      "ci_high": "0.554322457378013",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16611295681063123",
      "ci_low": "0.125",
      "ci_high": "0.21017355682031388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.21262458471760798",
      "ci_low": "0.167192429022082",
      "ci_high": "0.2607260726072607",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.78125",
      "ci_low": "0.6727272727272727",
      "ci_high": "0.8787878787878788",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.046511627906976744",
      "ci_low": "0.024052923315915836",
      "ci_high": "0.07165134402873755",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16279069767441862",
      "ci_low": "0.12244897959183673",
      "ci_high": "0.20569620253164558",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "27.774086378737543",
      "ci_low": "23.926013961388684",
      "ci_high": "31.83334775086505",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "11.49501661129568",
      "ci_low": "5.460672759877365",
      "ci_high": "17.45966998980028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16558021589202485",
      "ci_low": "0.12916420356021197",
      "ci_high": "0.1882910338776659",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.17385922668662604",
      "ci_low": "0.13562241373822254",
      "ci_high": "0.19770558557154919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.0641025641025641",
      "ci_low": "0.0136986301369863",
      "ci_high": "0.12658227848101267",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.0641025641025641",
      "ci_low": "0.0136986301369863",
      "ci_high": "0.12658227848101267",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "9945",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.0641025641025641",
      "ci_low": "0.013509009009009015",
      "ci_high": "0.125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "19.615384615384613",
      "ci_low": "14.202764358561462",
      "ci_high": "25.774647887323944",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "13.205128205128204",
      "ci_low": "4.84375",
      "ci_high": "21.566394148020652",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.11249999999999999",
      "ci_low": "0.020833333333333332",
      "ci_high": "0.15416666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.11249999999999999",
      "ci_low": "0.020833333333333332",
      "ci_high": "0.15416666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.10802469135802469",
      "ci_low": "0.0732484076433121",
      "ci_high": "0.14679158243922663",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.18209876543209877",
      "ci_low": "0.1387057909055577",
      "ci_high": "0.22769306930693067",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5932203389830508",
      "ci_low": "0.4528301886792453",
      "ci_high": "0.7288135593220338",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.07407407407407407",
      "ci_low": "0.045317220543806644",
      "ci_high": "0.10632183908045977",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.13271604938271606",
      "ci_low": "0.09523444347063979",
      "ci_high": "0.17365269461077845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "19.62962962962963",
      "ci_low": "16.149249096134234",
      "ci_high": "23.36449332863105",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "6.358024691358025",
      "ci_low": "1.7764578877275112",
      "ci_high": "11.008878496609084",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.15916400672498232",
      "ci_low": "0.10907171247539746",
      "ci_high": "0.18310944622715564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.17172958620327042",
      "ci_low": "0.11768263714450776",
      "ci_high": "0.19756545513982582",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.18421052631578946",
      "ci_low": "0.09722222222222222",
      "ci_high": "0.2835820895522388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2894736842105263",
      "ci_low": "0.18643082380764683",
      "ci_high": "0.4025974025974026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6363636363636364",
      "ci_low": "0.4166666666666667",
      "ci_high": "0.8421052631578947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.10526315789473684",
      "ci_low": "0.04054054054054054",
      "ci_high": "0.17978909559771578",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2631578947368421",
      "ci_low": "0.15625",
      "ci_high": "0.3780487804878049",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "32.36842105263158",
      "ci_low": "23.484461966604822",
      "ci_high": "41.39280620431449",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "6.052631578947368",
      "ci_low": "-6.765878090366581",
      "ci_high": "19.00035211267605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.14805194805194805",
      "ci_low": "0.08267274392274392",
      "ci_high": "0.18331414797323886",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.16285714285714287",
      "ci_low": "0.09094001831501831",
      "ci_high": "0.2016455627705628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5333333333333333",
      "ci_low": "0.43636363636363634",
      "ci_high": "0.6296296296296297",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.5081967213114754",
      "ci_high": "0.6907263355201498",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8888888888888888",
      "ci_low": "0.8059597844112769",
      "ci_high": "0.9622641509433962",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.06666666666666667",
      "ci_low": "0.021505376344086023",
      "ci_high": "0.11764705882352941",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.13333333333333333",
      "ci_low": "0.07272727272727272",
      "ci_high": "0.2",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "65.23809523809524",
      "ci_low": "57.46800920598389",
      "ci_high": "72.83357843137254",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "51.904761904761905",
      "ci_low": "39.36825307950728",
      "ci_high": "63.571592398427256",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5720583860583861",
      "ci_low": "0.4137630670771644",
      "ci_high": "0.6001817644229408",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5958941521441522",
      "ci_low": "0.4310031948720463",
      "ci_high": "0.6251893379405634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.41754385964912283",
      "ci_low": "0.35815602836879434",
      "ci_high": "0.48",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.45964912280701753",
      "ci_low": "0.39855072463768115",
      "ci_high": "0.5226480836236934",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.9083969465648855",
      "ci_low": "0.847995652173913",
      "ci_high": "0.9600066225165561",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.042105263157894736",
      "ci_low": "0.01818181818181818",
      "ci_high": "0.07063197026022305",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.17543859649122806",
      "ci_low": "0.1310344827586207",
      "ci_high": "0.22183098591549297",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "54.73684210526316",
      "ci_low": "49.82392810929179",
      "ci_high": "59.79094076655052",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "37.19298245614035",
      "ci_low": "29.543567243548978",
      "ci_high": "44.84538721368003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4632091458681123",
      "ci_low": "0.39594358076506175",
      "ci_high": "0.48295268414449993",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4632091458681123",
      "ci_low": "0.39594358076506175",
      "ci_high": "0.48295268414449993",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.2653061224489796",
      "ci_low": "0.20832908163265307",
      "ci_high": "0.3254014396456256",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.3551020408163265",
      "ci_low": "0.2923680860710051",
      "ci_high": "0.4184123458610583",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7471264367816092",
      "ci_low": "0.6451612903225806",
      "ci_high": "0.845360824742268",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.08979591836734693",
      "ci_low": "0.052862988431646434",
      "ci_high": "0.13008130081300814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.12653061224489795",
      "ci_low": "0.08366533864541832",
      "ci_high": "0.17269076305220885",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "36.44897959183674",
      "ci_low": "31.15380184331797",
      "ci_high": "42.0297416460396",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "23.79591836734694",
      "ci_low": "16.99989270386266",
      "ci_high": "30.763277374202687",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.34928537071394217",
      "ci_low": "0.2458757112422526",
      "ci_high": "0.34844482237339375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.3729657348301416",
      "ci_low": "0.26254525098749004",
      "ci_high": "0.3720682001614205",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.47692307692307695",
      "ci_low": "0.39090198863636366",
      "ci_high": "0.5666666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5153846153846153",
      "ci_low": "0.4271715210355987",
      "ci_high": "0.6043205733920326",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.9253731343283582",
      "ci_low": "0.8571428571428571",
      "ci_high": "0.9838709677419355",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.038461538461538464",
      "ci_low": "0.008064516129032258",
      "ci_high": "0.07518796992481203",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.14615384615384616",
      "ci_low": "0.08526719445818902",
      "ci_high": "0.21311475409836064",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "60.230769230769226",
      "ci_low": "53.27731092436974",
      "ci_high": "67.35320855614974",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "45.61538461538461",
      "ci_low": "34.869202898550725",
      "ci_high": "55.85229468599034",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5118422318422318",
      "ci_low": "0.3574269680936347",
      "ci_high": "0.5041812169312169",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5234750098386461",
      "ci_low": "0.365550308277581",
      "ci_high": "0.5156398809523809",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.11842105263157894",
      "ci_low": "0.05063291139240506",
      "ci_high": "0.19480519480519481",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.15789473684210525",
      "ci_low": "0.08",
      "ci_high": "0.24324324324324326",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.75",
      "ci_low": "0.4666666666666667",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.039473684210526314",
      "ci_low": "0.0",
      "ci_high": "0.08823529411764706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.17105263157894737",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.25758797653958926",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "19.473684210526315",
      "ci_low": "12.432432432432432",
      "ci_high": "27.027027027027025",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "2.3684210526315788",
      "ci_low": "-9.078947368421053",
      "ci_high": "13.733373493975902",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.18055555555555555",
      "ci_low": "0.0625",
      "ci_high": "0.16666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.18840579710144925",
      "ci_low": "0.06521739130434782",
      "ci_high": "0.17391304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.30612244897959184",
      "ci_low": "0.21176470588235294",
      "ci_high": "0.40426129572077446",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.32653061224489793",
      "ci_low": "0.23076923076923078",
      "ci_high": "0.425531914893617",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.9375",
      "ci_low": "0.8387096774193549",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.02040816326530612",
      "ci_low": "0.0",
      "ci_high": "0.05263157894736842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.07692307692307693",
      "ci_high": "0.21505376344086022",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "41.224489795918366",
      "ci_low": "32.825659501709815",
      "ci_high": "49.909988249118676",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "26.93877551020408",
      "ci_low": "14.807500000000001",
      "ci_high": "38.51083530338849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2932712215320911",
      "ci_low": "0.16500862663906143",
      "ci_high": "0.2916230176828003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2932712215320911",
      "ci_low": "0.16500862663906143",
      "ci_high": "0.2916230176828003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.1588785046728972",
      "ci_low": "0.09174120795107034",
      "ci_high": "0.23232323232323232",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.21495327102803738",
      "ci_low": "0.14",
      "ci_high": "0.29565882874889077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7391304347826086",
      "ci_low": "0.5416666666666666",
      "ci_high": "0.9090909090909091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.056074766355140186",
      "ci_low": "0.018018018018018018",
      "ci_high": "0.10281096444744184",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.12149532710280374",
      "ci_low": "0.06306306306306306",
      "ci_high": "0.18691588785046728",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "26.074766355140188",
      "ci_low": "19.428571428571427",
      "ci_high": "33.11940940366972",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "13.925233644859814",
      "ci_low": "4.68049645390071",
      "ci_high": "22.869585346215782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.18214285714285713",
      "ci_low": "0.08760416666666666",
      "ci_high": "0.18956043956043958",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.18571428571428572",
      "ci_low": "0.0893218954248366",
      "ci_high": "0.19327731092436973",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3082706766917293",
      "ci_low": "0.22949965846994536",
      "ci_high": "0.3902480788506515",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.40601503759398494",
      "ci_low": "0.32025490196078427",
      "ci_high": "0.49296019170579025",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.7592592592592593",
      "ci_low": "0.6271123197571464",
      "ci_high": "0.8805970149253731",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.09774436090225563",
      "ci_low": "0.045454545454545456",
      "ci_high": "0.15909452808338154",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.2781954887218045",
      "ci_low": "0.2",
      "ci_high": "0.358974358974359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "46.54135338345864",
      "ci_low": "39.91930485203945",
      "ci_high": "53.44619435612083",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "18.721804511278194",
      "ci_low": "7.5608401084010834",
      "ci_high": "30.138242157116924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.2878095238095238",
      "ci_low": "0.19799642857142855",
      "ci_high": "0.2964254604872252",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.2936831875607386",
      "ci_low": "0.20203717201166183",
      "ci_high": "0.30247495968084204",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "0",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "SpeciesNet v4.0.3b / operational_country_ablation",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5384615384615384",
      "ci_low": "0.48853078890626983",
      "ci_high": "0.5871559633027523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6806526806526807",
      "ci_low": "0.6353699270862545",
      "ci_high": "0.7245667370567246",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.791095890410959",
      "ci_low": "0.743201123618756",
      "ci_high": "0.8371215205803968",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.14219114219114218",
      "ci_low": "0.1100244498777506",
      "ci_high": "0.17592875577288777",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.21445221445221446",
      "ci_low": "0.1755536858974359",
      "ci_high": "0.25547772577002387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "67.6923076923077",
      "ci_low": "64.03218871364032",
      "ci_high": "71.298427928904",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "46.247086247086244",
      "ci_low": "39.473531920077974",
      "ci_high": "52.86408066642694",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4429138522252915",
      "ci_low": "0.37837072021640095",
      "ci_high": "0.45170270382921246",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5643579729967425",
      "ci_low": "0.4821175305983174",
      "ci_high": "0.57555667100819",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3518005540166205",
      "ci_low": "0.30057762875884914",
      "ci_high": "0.40298507462686567",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5263157894736842",
      "ci_low": "0.47246156617823687",
      "ci_high": "0.5795454545454546",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.6684210526315789",
      "ci_low": "0.599009900990099",
      "ci_high": "0.736559795436664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.1745152354570637",
      "ci_low": "0.1352112676056338",
      "ci_high": "0.21487603305785125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.2853185595567867",
      "ci_low": "0.2371967654986523",
      "ci_high": "0.33246897163120537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "54.958448753462605",
      "ci_low": "50.96579982517483",
      "ci_high": "59.05918062733916",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "26.426592797783933",
      "ci_low": "19.245597007859104",
      "ci_high": "33.78751613365839",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.28475413397310384",
      "ci_low": "0.21815943209231833",
      "ci_high": "0.29202158641477943",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3716282765411694",
      "ci_low": "0.28471654696794085",
      "ci_high": "0.3811129178633562",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6293333333333333",
      "ci_low": "0.5785090462363189",
      "ci_high": "0.680628272251309",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7893333333333333",
      "ci_low": "0.745945945945946",
      "ci_high": "0.8310258742367175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7972972972972973",
      "ci_low": "0.7507984700819382",
      "ci_high": "0.8421094139133322",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.16",
      "ci_low": "0.12403049641038227",
      "ci_high": "0.19726027397260273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.21866666666666668",
      "ci_low": "0.17595123712403662",
      "ci_high": "0.26285787923416787",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "78.32",
      "ci_low": "75.12817258883248",
      "ci_high": "81.45210546547388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "56.45333333333333",
      "ci_low": "49.41324782785467",
      "ci_high": "63.288444891725845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5094660158435669",
      "ci_low": "0.4416274509778075",
      "ci_high": "0.5124802425132557",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6400983275983275",
      "ci_low": "0.554865258920835",
      "ci_high": "0.64388543290127",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.12643678160919541",
      "ci_low": "0.07975460122699386",
      "ci_high": "0.1783783783783784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.26436781609195403",
      "ci_low": "0.2010575192276223",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.4782608695652174",
      "ci_low": "0.3333333333333333",
      "ci_high": "0.6222222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.13793103448275862",
      "ci_low": "0.08928571428571429",
      "ci_high": "0.19047619047619047",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.27586206896551724",
      "ci_low": "0.21052631578947367",
      "ci_high": "0.3423913043478261",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "28.218390804597703",
      "ci_low": "23.394955300127716",
      "ci_high": "33.260898628225995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.632183908045977",
      "ci_low": "-8.04762443438914",
      "ci_high": "9.285714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.12717480931766648",
      "ci_low": "0.06375661375661375",
      "ci_high": "0.12991307634164775",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1387361556192725",
      "ci_low": "0.06955266955266956",
      "ci_high": "0.1417233560090703",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.4125",
      "ci_low": "0.35019455252918286",
      "ci_high": "0.4756583180715735",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5791666666666667",
      "ci_low": "0.5145203279604212",
      "ci_high": "0.6428571428571429",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.7122302158273381",
      "ci_low": "0.6381459763539283",
      "ci_high": "0.7862595419847328",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.16666666666666666",
      "ci_low": "0.12083333333333333",
      "ci_high": "0.21428571428571427",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2708333333333333",
      "ci_low": "0.21518987341772153",
      "ci_high": "0.32644628099173556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "60.41666666666667",
      "ci_low": "55.72071117727897",
      "ci_high": "65.13274336283186",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "33.333333333333336",
      "ci_low": "24.602341315766047",
      "ci_high": "41.96760175240249",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.4520707900018244",
      "ci_low": "0.32700141493244944",
      "ci_high": "0.4337621462190427",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.49785011050833833",
      "ci_low": "0.36011548226738105",
      "ci_high": "0.47768742684881915",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.44801980198019803",
      "ci_high": "0.551219512195122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.665",
      "ci_low": "0.6159600997506235",
      "ci_high": "0.7113924050632912",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7518796992481203",
      "ci_low": "0.6966280130605974",
      "ci_high": "0.8042704626334519",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.165",
      "ci_low": "0.1287122076818942",
      "ci_high": "0.2034313725490196",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.245",
      "ci_low": "0.20191587193460492",
      "ci_high": "0.28967372113279505",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "64.475",
      "ci_low": "60.47858942065491",
      "ci_high": "68.35655381944444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "39.975",
      "ci_low": "32.46408831640229",
      "ci_high": "47.203308410737975",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.3868238089527153",
      "ci_low": "0.3507570427397413",
      "ci_high": "0.4062130086997508",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.560894522981437",
      "ci_low": "0.5085977119726248",
      "ci_high": "0.5890088626146385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.40512820512820513",
      "ci_low": "0.35426873953098825",
      "ci_high": "0.45641439205955325",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5538461538461539",
      "ci_low": "0.5013546502757252",
      "ci_high": "0.6058201058201058",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.7314814814814815",
      "ci_low": "0.6698113207547169",
      "ci_high": "0.7903950925348304",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.14871794871794872",
      "ci_low": "0.11398963730569948",
      "ci_high": "0.18414322250639387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.24871794871794872",
      "ci_low": "0.2040289735203078",
      "ci_high": "0.2948717948717949",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "59.205128205128204",
      "ci_low": "55.328277857313296",
      "ci_high": "63.10917372521893",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "34.33333333333333",
      "ci_low": "27.27512896825397",
      "ci_high": "41.21699574138598",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.37517134480551423",
      "ci_low": "0.33174762079806003",
      "ci_high": "0.39839433383733197",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.47136912552487686",
      "ci_low": "0.4168111133103831",
      "ci_high": "0.5005467271289555",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6245847176079734",
      "ci_low": "0.5691305106617025",
      "ci_high": "0.6786885245901639",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7574750830564784",
      "ci_low": "0.7079037800687286",
      "ci_high": "0.8042704626334519",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8245614035087719",
      "ci_low": "0.7735828898564748",
      "ci_high": "0.8724279835390947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.132890365448505",
      "ci_low": "0.0958904109589041",
      "ci_high": "0.17288322140496745",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.18272425249169436",
      "ci_low": "0.14012738853503184",
      "ci_high": "0.22772277227722773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "72.42524916943522",
      "ci_low": "68.11415853449752",
      "ci_high": "76.61260894444932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "54.152823920265774",
      "ci_low": "46.591564593661616",
      "ci_high": "61.53155395900321",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5270296705358626",
      "ci_low": "0.45741056628735766",
      "ci_high": "0.5469665253068151",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6587870881698281",
      "ci_low": "0.5717632078591971",
      "ci_high": "0.6837081566335188",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.14736638106895145",
      "ci_high": "0.2549019607843137",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.362962962962963",
      "ci_low": "0.3036419172932331",
      "ci_high": "0.42231255483481234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.5510204081632653",
      "ci_low": "0.44230263157894734",
      "ci_high": "0.6588302139037432",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.16296296296296298",
      "ci_low": "0.11904761904761904",
      "ci_high": "0.20784313725490197",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.35185185185185186",
      "ci_low": "0.2886139774859287",
      "ci_high": "0.4144486692015209",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "45.48148148148148",
      "ci_low": "41.07802496385791",
      "ci_high": "50.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "10.296296296296298",
      "ci_low": "1.512737038567297",
      "ci_high": "19.3106822923058",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.15882654783583577",
      "ci_low": "0.1257856911894179",
      "ci_high": "0.1799348546313385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2470635188557445",
      "ci_low": "0.19566663073909452",
      "ci_high": "0.27989866275985986",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5296803652968036",
      "ci_low": "0.461856258148631",
      "ci_high": "0.5968586387434555",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7123287671232876",
      "ci_low": "0.6485120258601202",
      "ci_high": "0.7721518987341772",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7435897435897436",
      "ci_low": "0.6713286713286714",
      "ci_high": "0.8128664402942841",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.182648401826484",
      "ci_low": "0.1318181818181818",
      "ci_high": "0.23660714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.2054794520547945",
      "ci_low": "0.1512168021680217",
      "ci_high": "0.26244343891402716",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "67.57990867579909",
      "ci_low": "62.37902247103038",
      "ci_high": "72.62233486943164",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "47.03196347031963",
      "ci_low": "38.10506520726595",
      "ci_high": "55.775018532246094",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5139975937904016",
      "ci_low": "0.4186375573303413",
      "ci_high": "0.5457569538817727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6074517017522928",
      "ci_low": "0.4947534768449488",
      "ci_high": "0.6449854909511858",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6232876712328768",
      "ci_low": "0.5663082437275986",
      "ci_high": "0.6784452296819788",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7534246575342466",
      "ci_low": "0.7035714285714286",
      "ci_high": "0.8013256373038192",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8272727272727273",
      "ci_low": "0.7757009345794392",
      "ci_high": "0.8756218905472637",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.13013698630136986",
      "ci_low": "0.09287870411322424",
      "ci_high": "0.17105483927940654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.1815068493150685",
      "ci_low": "0.1378202296819788",
      "ci_high": "0.2264808362369338",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "72.1917808219178",
      "ci_low": "67.77777777777777",
      "ci_high": "76.48747617506383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "54.04109589041096",
      "ci_low": "46.33561643835616",
      "ci_high": "61.63704104065199",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.5348133848452552",
      "ci_low": "0.45390253636973366",
      "ci_high": "0.5514384324631775",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6652556738319029",
      "ci_low": "0.5646104720696687",
      "ci_high": "0.6859356111127329",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5421052631578948",
      "ci_low": "0.465435506684309",
      "ci_high": "0.6171479591836734",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7157894736842105",
      "ci_low": "0.6412973801560757",
      "ci_high": "0.7853658536585366",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7573529411764706",
      "ci_low": "0.6796875",
      "ci_high": "0.8306451612903226",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.1736842105263158",
      "ci_low": "0.11904761904761904",
      "ci_high": "0.233502538071066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.19473684210526315",
      "ci_low": "0.1375630030843301",
      "ci_high": "0.2561576354679803",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "68.68421052631578",
      "ci_low": "62.70930517377459",
      "ci_high": "74.46080316742082",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "49.21052631578948",
      "ci_low": "39.371720269259534",
      "ci_high": "58.54855025172478",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.41252119439710266",
      "ci_low": "0.3256597225503817",
      "ci_high": "0.4342678355843679",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.49301313476726905",
      "ci_low": "0.3892030830480171",
      "ci_high": "0.5190030230154641",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2247191011235955",
      "ci_low": "0.1410211267605634",
      "ci_high": "0.31429373996789717",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.39325842696629215",
      "ci_low": "0.2923076923076923",
      "ci_high": "0.4948453608247423",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.5714285714285714",
      "ci_low": "0.40540540540540543",
      "ci_high": "0.7333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.16853932584269662",
      "ci_low": "0.09375",
      "ci_high": "0.25",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.29213483146067415",
      "ci_low": "0.19468411996066864",
      "ci_high": "0.4024390243902439",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "46.06741573033708",
      "ci_low": "38.872282237111996",
      "ci_high": "53.146067415730336",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "16.85393258426966",
      "ci_low": "1.904642857142857",
      "ci_high": "30.32967032967033",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.1830952380952381",
      "ci_low": "0.08504059829059828",
      "ci_high": "0.20305584554334558",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2496753246753247",
      "ci_low": "0.1159644522144522",
      "ci_high": "0.27689433483183484",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.07692307692307693",
      "ci_high": "0.6666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.75",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "9817",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.08333333333333333",
      "ci_high": "0.6",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "45.0",
      "ci_low": "23.333333333333336",
      "ci_high": "70.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "11.666666666666668",
      "ci_low": "-27.5",
      "ci_high": "55.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3",
      "ci_low": "0.0",
      "ci_high": "0.3",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.22885572139303484",
      "ci_low": "0.16922885572139304",
      "ci_high": "0.29223919915700736",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.4079601990049751",
      "ci_low": "0.34146341463414637",
      "ci_high": "0.475",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5609756097560976",
      "ci_low": "0.4487068965517242",
      "ci_high": "0.6714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.1791044776119403",
      "ci_low": "0.12735849056603774",
      "ci_high": "0.23113266805726745",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3681592039800995",
      "ci_low": "0.295451601507301",
      "ci_high": "0.44",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "48.00995024875622",
      "ci_low": "43.24600254704967",
      "ci_high": "52.9145054945055",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "11.194029850746269",
      "ci_low": "0.9934554973822001",
      "ci_high": "21.682494608195537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.20066466129814545",
      "ci_low": "0.1432497857253751",
      "ci_high": "0.2229916581867939",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.28984895520843235",
      "ci_low": "0.2069163571588752",
      "ci_high": "0.322099061825369",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "-80.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "70.0",
      "ci_low": "40.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "70.0",
      "ci_low": "40.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2857142857142857",
      "ci_low": "0.17543859649122806",
      "ci_high": "0.4",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.4714285714285714",
      "ci_low": "0.3492063492063492",
      "ci_high": "0.5890472113502935",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6060606060606061",
      "ci_low": "0.42105263157894735",
      "ci_high": "0.78125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.18571428571428572",
      "ci_low": "0.0945945945945946",
      "ci_high": "0.2835820895522388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.22857142857142856",
      "ci_low": "0.12902073732718894",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "48.285714285714285",
      "ci_low": "39.0",
      "ci_high": "57.28454415954415",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "25.428571428571427",
      "ci_low": "9.666666666666666",
      "ci_high": "40.62524801587301",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.26960917870008777",
      "ci_low": "0.16741883116883113",
      "ci_high": "0.3370668468356888",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.42367156652870935",
      "ci_low": "0.2630867346938775",
      "ci_high": "0.5296764735989395",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6444444444444445",
      "ci_low": "0.5",
      "ci_high": "0.7857435597189691",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8",
      "ci_low": "0.675",
      "ci_high": "0.9130434782608695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8055555555555556",
      "ci_low": "0.6666666666666666",
      "ci_high": "0.9411764705882353",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.15555555555555556",
      "ci_low": "0.046511627906976744",
      "ci_high": "0.2777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.17777777777777778",
      "ci_low": "0.0625",
      "ci_high": "0.30434782608695654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "74.88888888888889",
      "ci_low": "63.499709302325584",
      "ci_high": "85.5813953488372",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "57.111111111111114",
      "ci_low": "37.65830249768733",
      "ci_high": "76.19047619047619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5516679399032339",
      "ci_low": "0.31077614379084967",
      "ci_high": "0.5687562656641603",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7214119214119215",
      "ci_low": "0.4063995726495726",
      "ci_high": "0.7437581935608252",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.18222222222222223",
      "ci_low": "0.12598425196850394",
      "ci_high": "0.24255732286717613",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3422222222222222",
      "ci_low": "0.276",
      "ci_high": "0.40816751700680265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5324675324675324",
      "ci_low": "0.4067664194915255",
      "ci_high": "0.656261660447761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.16",
      "ci_low": "0.11160714285714286",
      "ci_high": "0.21032268170426058",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.36",
      "ci_low": "0.291891245312298",
      "ci_high": "0.42857142857142855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "44.44444444444444",
      "ci_low": "39.67722915398661",
      "ci_high": "49.32447514478764",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "8.444444444444445",
      "ci_low": "-1.1556018518518516",
      "ci_high": "18.550676877470348",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.16449314574314572",
      "ci_low": "0.11707052669552669",
      "ci_high": "0.18846242360580595",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.24369354924910483",
      "ci_low": "0.1734378173267062",
      "ci_high": "0.2792035905271199",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4225352112676056",
      "ci_low": "0.3055555555555556",
      "ci_high": "0.5396825396825397",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.4915179898899792",
      "ci_high": "0.7192982456140351",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6976744186046512",
      "ci_low": "0.5526315789473685",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.18309859154929578",
      "ci_low": "0.09721102150537636",
      "ci_high": "0.2753623188405797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.2676056338028169",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.37333333333333335",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "59.71830985915493",
      "ci_low": "50.61538461538461",
      "ci_high": "68.52950980392157",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "32.95774647887324",
      "ci_low": "16.51497113997114",
      "ci_high": "48.80014925373134",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3590117767537122",
      "ci_low": "0.22362350463156913",
      "ci_high": "0.38556098866345545",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4121987066431511",
      "ci_low": "0.25675291272513495",
      "ci_high": "0.44268113513211554",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6245847176079734",
      "ci_low": "0.5691305106617025",
      "ci_high": "0.6786885245901639",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7574750830564784",
      "ci_low": "0.7079037800687286",
      "ci_high": "0.8042704626334519",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8245614035087719",
      "ci_low": "0.7735828898564748",
      "ci_high": "0.8724279835390947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.132890365448505",
      "ci_low": "0.0958904109589041",
      "ci_high": "0.17288322140496745",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.18272425249169436",
      "ci_low": "0.14012738853503184",
      "ci_high": "0.22772277227722773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "72.42524916943522",
      "ci_low": "68.11415853449752",
      "ci_high": "76.61260894444932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "54.152823920265774",
      "ci_low": "46.591564593661616",
      "ci_high": "61.53155395900321",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5270296705358626",
      "ci_low": "0.45741056628735766",
      "ci_high": "0.5469665253068151",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6587870881698281",
      "ci_low": "0.5717632078591971",
      "ci_high": "0.6837081566335188",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6410256410256411",
      "ci_low": "0.525",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8333333333333334",
      "ci_low": "0.7368421052631579",
      "ci_high": "0.9195402298850575",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7692307692307693",
      "ci_low": "0.6567164179104478",
      "ci_high": "0.8656716417910447",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.19230769230769232",
      "ci_low": "0.10989010989010989",
      "ci_high": "0.2894736842105263",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.20512820512820512",
      "ci_low": "0.12161138411138413",
      "ci_high": "0.30434782608695654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "78.07692307692308",
      "ci_low": "70.43478260869566",
      "ci_high": "85.13157894736842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "57.56410256410256",
      "ci_low": "42.20522247360483",
      "ci_high": "71.0962709284627",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.46781204906204904",
      "ci_low": "0.3470598580341228",
      "ci_high": "0.5025694510758849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7017180735930736",
      "ci_low": "0.5205897870511842",
      "ci_high": "0.7538541766138273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5493827160493827",
      "ci_low": "0.4939015038156516",
      "ci_high": "0.6046511627906976",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.6975308641975309",
      "ci_low": "0.6470588235294118",
      "ci_high": "0.7461785249036835",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7876106194690266",
      "ci_low": "0.7324561403508771",
      "ci_high": "0.8407079646017699",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.14814814814814814",
      "ci_low": "0.11038961038961038",
      "ci_high": "0.18688601195036003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.20987654320987653",
      "ci_low": "0.1651376146788991",
      "ci_high": "0.256797583081571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "67.8395061728395",
      "ci_low": "63.52756746693145",
      "ci_high": "72.04346488119661",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "46.85185185185185",
      "ci_low": "38.94552957691935",
      "ci_high": "54.43718835216206",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.4111462295345362",
      "ci_low": "0.37239530279133876",
      "ci_high": "0.4320253382883027",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5626211562051548",
      "ci_low": "0.5095935722407794",
      "ci_high": "0.5911925681839932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2894736842105263",
      "ci_low": "0.18292682926829268",
      "ci_high": "0.4024429838454228",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.5263157894736842",
      "ci_low": "0.4074074074074074",
      "ci_high": "0.6406350160256409",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.55",
      "ci_low": "0.3870967741935484",
      "ci_high": "0.7058823529411765",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.33784259611724393",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.39473684210526316",
      "ci_low": "0.2816791373239437",
      "ci_high": "0.5087884806355509",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "50.131578947368425",
      "ci_low": "41.95395290159798",
      "ci_high": "58.52471070395371",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "10.657894736842106",
      "ci_low": "-6.7089963833634725",
      "ci_high": "28.219580983078156",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.21617186059376728",
      "ci_low": "0.10451746529332735",
      "ci_high": "0.24018622182415286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.31344919786096254",
      "ci_low": "0.1515503246753247",
      "ci_high": "0.34827002164502163",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5047619047619047",
      "ci_low": "0.39795644064077246",
      "ci_high": "0.6100097402597401",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6285714285714286",
      "ci_low": "0.5306122448979592",
      "ci_high": "0.7265151515151512",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.803030303030303",
      "ci_low": "0.6956521739130435",
      "ci_high": "0.8947368421052632",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.12380952380952381",
      "ci_low": "0.0660377358490566",
      "ci_high": "0.18919630156472253",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.22857142857142856",
      "ci_low": "0.14034677816035415",
      "ci_high": "0.32075471698113206",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "67.23809523809524",
      "ci_low": "59.77528089887641",
      "ci_high": "74.60954968944098",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "44.38095238095238",
      "ci_low": "29.82861635220126",
      "ci_high": "58.285778635778634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5040959403704502",
      "ci_low": "0.348370453769799",
      "ci_high": "0.5282053308562971",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5671079329167564",
      "ci_low": "0.39191676049102386",
      "ci_high": "0.5942309972133342",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3684210526315789",
      "ci_low": "0.3103448275862069",
      "ci_high": "0.42758620689655175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5263157894736842",
      "ci_low": "0.4659854802135908",
      "ci_high": "0.5860927152317881",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.7",
      "ci_low": "0.625",
      "ci_high": "0.7730496453900709",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.15789473684210525",
      "ci_low": "0.11610079791564892",
      "ci_high": "0.20074384645713642",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.256140350877193",
      "ci_low": "0.20512820512820512",
      "ci_high": "0.30827577418121566",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "56.24561403508772",
      "ci_low": "51.620792470737115",
      "ci_high": "60.87842596117244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "30.631578947368425",
      "ci_low": "22.666563467492256",
      "ci_high": "38.398612099644126",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.32005562008879124",
      "ci_low": "0.256300135050008",
      "ci_high": "0.33979984290336984",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.40212116370130185",
      "ci_low": "0.3220181183961639",
      "ci_high": "0.4269280077503878",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.689795918367347",
      "ci_low": "0.6282051282051282",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8448979591836735",
      "ci_low": "0.7972350230414746",
      "ci_high": "0.8906882591093117",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8164251207729468",
      "ci_low": "0.7625539241655945",
      "ci_high": "0.8679245283018868",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.15510204081632653",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.2008201008299609",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.19183673469387755",
      "ci_low": "0.1405605894652992",
      "ci_high": "0.2446351931330472",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "82.16326530612244",
      "ci_low": "78.4765625",
      "ci_high": "85.73532261741907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "62.97959183673469",
      "ci_low": "54.62435938232733",
      "ci_high": "71.02882171387074",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.5430831128747795",
      "ci_low": "0.44114724877745715",
      "ci_high": "0.5395518275864477",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.662745493677697",
      "ci_low": "0.5383491849487613",
      "ci_high": "0.6584361285800718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5153846153846153",
      "ci_low": "0.4295774647887324",
      "ci_high": "0.6050420168067226",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6846153846153846",
      "ci_low": "0.6027397260273972",
      "ci_high": "0.7647058823529411",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7528089887640449",
      "ci_low": "0.6623332748332749",
      "ci_high": "0.8421052631578947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.16923076923076924",
      "ci_low": "0.10686240770867227",
      "ci_high": "0.23448803659394785",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.2692307692307692",
      "ci_low": "0.19117412888722368",
      "ci_high": "0.34814814814814815",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "71.07692307692307",
      "ci_low": "65.40700483091787",
      "ci_high": "76.79256338443396",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "44.15384615384616",
      "ci_low": "31.904386951631047",
      "ci_high": "56.35594862698725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.42257495590828925",
      "ci_low": "0.30799402757736094",
      "ci_high": "0.4257711372989151",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5186147186147186",
      "ci_low": "0.37799267020857935",
      "ci_high": "0.5225373048668502",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.09210526315789473",
      "ci_low": "0.03076923076923077",
      "ci_high": "0.16304347826086957",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.21052631578947367",
      "ci_low": "0.12194297956493079",
      "ci_high": "0.3055555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.4375",
      "ci_low": "0.19047619047619047",
      "ci_high": "0.6923076923076923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.11842105263157894",
      "ci_low": "0.05063291139240506",
      "ci_high": "0.1969750320102432",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.2631578947368421",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.36619718309859156",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "25.789473684210527",
      "ci_low": "19.23051948051948",
      "ci_high": "32.706092436974785",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "-0.5263157894736842",
      "ci_low": "-13.7102688172043",
      "ci_high": "12.428667953667953",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.10690943043884221",
      "ci_low": "0.027450980392156862",
      "ci_high": "0.1134453781512605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.11853002070393373",
      "ci_low": "0.03043478260869565",
      "ci_high": "0.12577639751552794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.15306122448979592",
      "ci_low": "0.08602150537634409",
      "ci_high": "0.22916666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.30612244897959184",
      "ci_low": "0.21739130434782608",
      "ci_high": "0.40404040404040403",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.32",
      "ci_high": "0.6785714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.15306122448979592",
      "ci_low": "0.08510136491368929",
      "ci_high": "0.2289156626506024",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2857142857142857",
      "ci_low": "0.2",
      "ci_high": "0.37777777777777777",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "30.10204081632653",
      "ci_low": "23.22908266129032",
      "ci_high": "37.44465408805031",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "1.5306122448979593",
      "ci_low": "-10.568262987012988",
      "ci_high": "13.57909632571996",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.12275641025641025",
      "ci_low": "0.06046932234432234",
      "ci_high": "0.14026556776556773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.13876811594202898",
      "ci_low": "0.06835662525879917",
      "ci_high": "0.1585610766045548",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5046728971962616",
      "ci_low": "0.4090909090909091",
      "ci_high": "0.5979381443298969",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6355140186915887",
      "ci_low": "0.5416666666666666",
      "ci_high": "0.7254901960784313",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7941176470588235",
      "ci_low": "0.6935483870967742",
      "ci_high": "0.8852459016393442",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.1308411214953271",
      "ci_low": "0.07079646017699115",
      "ci_high": "0.2",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2336448598130841",
      "ci_low": "0.15730337078651685",
      "ci_high": "0.315219839404622",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "64.01869158878505",
      "ci_low": "56.4925045823514",
      "ci_high": "71.19666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "40.65420560747664",
      "ci_low": "27.430896715004444",
      "ci_high": "53.39632075471699",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5196428571428571",
      "ci_low": "0.3259757653061225",
      "ci_high": "0.4816921768707482",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5705882352941177",
      "ci_low": "0.3579341736694677",
      "ci_high": "0.5289169000933706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3383458646616541",
      "ci_low": "0.25862068965517243",
      "ci_high": "0.4198473282442748",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5338345864661654",
      "ci_low": "0.4473684210526316",
      "ci_high": "0.6197183098591549",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6338028169014085",
      "ci_low": "0.5185185185185185",
      "ci_high": "0.746031746031746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.19548872180451127",
      "ci_low": "0.12799548872180452",
      "ci_high": "0.26515151515151514",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3007518796992481",
      "ci_low": "0.2222222222222222",
      "ci_high": "0.3800041322314049",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "57.51879699248121",
      "ci_low": "51.55024579315561",
      "ci_high": "63.593906249999996",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "27.44360902255639",
      "ci_low": "15.867741046831956",
      "ci_high": "38.87218045112782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3334750566893424",
      "ci_low": "0.20898224914965985",
      "ci_high": "0.3308555581323438",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.38111435050210557",
      "ci_low": "0.23883685617103984",
      "ci_high": "0.3781206378655358",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MINIMAL corrected",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5501165501165501",
      "ci_low": "0.5",
      "ci_high": "0.598086689967989",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6993006993006993",
      "ci_low": "0.6536547924689774",
      "ci_high": "0.743878874905425",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7866666666666666",
      "ci_low": "0.738093336376084",
      "ci_high": "0.8327759197324415",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.14918414918414918",
      "ci_low": "0.11566265060240964",
      "ci_high": "0.1848341232227488",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.2191142191142191",
      "ci_low": "0.18013747359434346",
      "ci_high": "0.26054884574224124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "69.06759906759906",
      "ci_low": "65.45428660178091",
      "ci_high": "72.62228346456692",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "47.15617715617715",
      "ci_low": "40.31235259433962",
      "ci_high": "53.75607613937594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.41656594752569986",
      "ci_low": "0.3575201952296675",
      "ci_high": "0.42504995181813016",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5643796708412706",
      "ci_low": "0.48438219998858173",
      "ci_high": "0.5758741282697247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3656509695290859",
      "ci_low": "0.3131868131868132",
      "ci_high": "0.418283937486606",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5429362880886427",
      "ci_low": "0.4888165452130524",
      "ci_high": "0.5966862119318952",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.673469387755102",
      "ci_low": "0.6038624555646706",
      "ci_high": "0.7417582417582418",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.1772853185595568",
      "ci_low": "0.13662790697674418",
      "ci_high": "0.21846448087431689",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.31301939058171746",
      "ci_low": "0.2631542898341745",
      "ci_high": "0.36193029490616624",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "55.98337950138504",
      "ci_low": "51.853715222043874",
      "ci_high": "60.03070436827523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "24.681440443213297",
      "ci_low": "17.08817226890756",
      "ci_high": "32.37504528985507",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.29717896697986373",
      "ci_low": "0.22896800683550192",
      "ci_high": "0.3051930029136122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3727329416357613",
      "ci_low": "0.28718021196317195",
      "ci_high": "0.38278444433232717",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6426666666666667",
      "ci_low": "0.5922077922077922",
      "ci_high": "0.6926832340111814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7973333333333333",
      "ci_low": "0.7542833200319234",
      "ci_high": "0.8388906926406926",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8060200668896321",
      "ci_low": "0.7593220338983051",
      "ci_high": "0.8491803278688524",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.15466666666666667",
      "ci_low": "0.11968028277868704",
      "ci_high": "0.19200263157894734",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.22933333333333333",
      "ci_low": "0.18701245908562983",
      "ci_high": "0.2727327624066754",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "78.8",
      "ci_low": "75.60900025197957",
      "ci_high": "81.9201843575419",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "55.86666666666667",
      "ci_low": "48.75688559322034",
      "ci_high": "62.80641137670538",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5238805312334723",
      "ci_low": "0.454377981352297",
      "ci_high": "0.5283396577396728",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6514924555082925",
      "ci_low": "0.5650597973227283",
      "ci_high": "0.6570377794967727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.16666666666666666",
      "ci_low": "0.11309271589991929",
      "ci_high": "0.2251655629139073",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2988505747126437",
      "ci_low": "0.23214285714285715",
      "ci_high": "0.3697954299132948",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5576923076923077",
      "ci_low": "0.42371251629726203",
      "ci_high": "0.693556616194865",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.13218390804597702",
      "ci_low": "0.08333333333333333",
      "ci_high": "0.18391157182145118",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.27011494252873564",
      "ci_low": "0.20555555555555555",
      "ci_high": "0.33689984887235525",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "31.32183908045977",
      "ci_low": "26.120154291224686",
      "ci_high": "36.72872451102788",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "4.310344827586206",
      "ci_low": "-4.890168970814131",
      "ci_high": "13.666954022988502",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1611013129531648",
      "ci_low": "0.08885768038545817",
      "ci_high": "0.16342025530751023",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.16947021232735518",
      "ci_low": "0.09347366378210534",
      "ci_high": "0.17190961921958867",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.4041666666666667",
      "ci_low": "0.34156378600823045",
      "ci_high": "0.4669421487603306",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.5377299528301887",
      "ci_high": "0.6625514403292181",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6736111111111112",
      "ci_low": "0.5961506622516556",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.19583333333333333",
      "ci_low": "0.14601532530669606",
      "ci_high": "0.24778857252789532",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.30833333333333335",
      "ci_low": "0.25",
      "ci_high": "0.36681222707423583",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "61.416666666666664",
      "ci_low": "56.85577441233857",
      "ci_high": "66.09248656544162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "30.583333333333332",
      "ci_low": "21.574670739027578",
      "ci_high": "39.641509517485616",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.41970945720945724",
      "ci_low": "0.30556971910665087",
      "ci_high": "0.40520989269568813",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.4675244586636991",
      "ci_low": "0.34038145925804153",
      "ci_high": "0.4513730450281083",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5075",
      "ci_low": "0.4556054298813513",
      "ci_high": "0.5578425920861516",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6825",
      "ci_low": "0.6346141975308641",
      "ci_high": "0.728",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7435897435897436",
      "ci_low": "0.6885783204684589",
      "ci_high": "0.7952231866921071",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.175",
      "ci_low": "0.13793103448275862",
      "ci_high": "0.21445783132530122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.255",
      "ci_low": "0.21148711403816428",
      "ci_high": "0.30025445292620867",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "65.325",
      "ci_low": "61.4003918336836",
      "ci_high": "69.10532186766127",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "39.825",
      "ci_low": "32.40685171292824",
      "ci_high": "47.04411679218715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.37121324271570644",
      "ci_low": "0.3358122897889391",
      "ci_high": "0.3900793684897925",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5661001951414523",
      "ci_low": "0.5121137419281322",
      "ci_high": "0.5948710369469336",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4230769230769231",
      "ci_low": "0.3700980392156863",
      "ci_high": "0.4755784061696658",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5717948717948718",
      "ci_low": "0.5187032418952618",
      "ci_high": "0.6246913580246913",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.7399103139013453",
      "ci_low": "0.6778242677824268",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.14871794871794872",
      "ci_low": "0.11315789473684211",
      "ci_high": "0.18666751918158567",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.2692307692307692",
      "ci_low": "0.225",
      "ci_high": "0.3153877005347593",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "60.794871794871796",
      "ci_low": "56.7909814504501",
      "ci_high": "64.74097301206173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "33.87179487179487",
      "ci_low": "26.572013956992848",
      "ci_high": "41.12089822024471",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.40579330224116467",
      "ci_low": "0.3570427674279906",
      "ci_high": "0.43146331137058513",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.48903295398294216",
      "ci_low": "0.4302823094645015",
      "ci_high": "0.519968606010705",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6345514950166113",
      "ci_low": "0.5798045602605864",
      "ci_high": "0.6883128002134852",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7807308970099668",
      "ci_low": "0.7331285877734598",
      "ci_high": "0.8269230769230769",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8127659574468085",
      "ci_low": "0.7614641144712501",
      "ci_high": "0.861244019138756",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.1461794019933555",
      "ci_low": "0.10726643598615918",
      "ci_high": "0.1875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.19601328903654486",
      "ci_low": "0.15202702702702703",
      "ci_high": "0.24138224255219326",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "73.62126245847176",
      "ci_low": "69.4196423608409",
      "ci_high": "77.77787258248009",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "54.01993355481728",
      "ci_low": "46.4026061045595",
      "ci_high": "61.45289307164307",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.53739400667129",
      "ci_low": "0.46643933526127884",
      "ci_high": "0.5587819969490437",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6448728080055479",
      "ci_low": "0.5597272023135346",
      "ci_high": "0.6705383963388526",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.23333333333333334",
      "ci_low": "0.1776061776061776",
      "ci_high": "0.2926852452935175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3851851851851852",
      "ci_low": "0.3228346456692913",
      "ci_high": "0.4461538461538462",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6057692307692307",
      "ci_low": "0.5046728971962616",
      "ci_high": "0.7043478260869566",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.15185185185185185",
      "ci_low": "0.11063153774994663",
      "ci_high": "0.195489324272219",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.362962962962963",
      "ci_low": "0.30370039682539685",
      "ci_high": "0.4221472510572856",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "48.18518518518519",
      "ci_low": "43.711182600311496",
      "ci_high": "52.811387900355875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "11.88888888888889",
      "ci_low": "2.985852797001154",
      "ci_high": "20.98177738469601",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.1932726823400198",
      "ci_low": "0.1530077659341675",
      "ci_high": "0.21777286574887691",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2863298997629923",
      "ci_low": "0.22667817175432228",
      "ci_high": "0.322626467776114",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5205479452054794",
      "ci_low": "0.4504088167325996",
      "ci_high": "0.5893719806763285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7168949771689498",
      "ci_low": "0.6528479418537708",
      "ci_high": "0.7767441860465116",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7261146496815286",
      "ci_low": "0.6490028309994439",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1963470319634703",
      "ci_low": "0.1415929203539823",
      "ci_high": "0.25490441176470585",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.228310502283105",
      "ci_low": "0.170303354083434",
      "ci_high": "0.2893434410900347",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "66.98630136986301",
      "ci_low": "61.84099897999541",
      "ci_high": "71.96263815986003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "44.15525114155251",
      "ci_low": "34.56017312782854",
      "ci_high": "53.27041987559242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4384750552365211",
      "ci_low": "0.36227901229237713",
      "ci_high": "0.46750781977760164",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.571346284096073",
      "ci_low": "0.4720605311688551",
      "ci_high": "0.6091768560738445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6301369863013698",
      "ci_low": "0.5743243243243243",
      "ci_high": "0.6847466838614591",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7773972602739726",
      "ci_low": "0.7286232076473712",
      "ci_high": "0.8245047635645405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8105726872246696",
      "ci_low": "0.7581395348837209",
      "ci_high": "0.8608695652173913",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.14726027397260275",
      "ci_low": "0.10702310283972405",
      "ci_high": "0.18900343642611683",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.19863013698630136",
      "ci_low": "0.15383986928104576",
      "ci_high": "0.24516129032258063",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "73.35616438356165",
      "ci_low": "69.03214605734767",
      "ci_high": "77.57901616787066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "53.49315068493151",
      "ci_low": "45.74319528519055",
      "ci_high": "61.082036676854685",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.5542891090104712",
      "ci_low": "0.47013065552770406",
      "ci_high": "0.5726016279496675",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6489238349390882",
      "ci_low": "0.5503968650080437",
      "ci_high": "0.6703628815020497",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5368421052631579",
      "ci_low": "0.45901639344262296",
      "ci_high": "0.6141304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7210526315789474",
      "ci_low": "0.6473684210526316",
      "ci_high": "0.7905759162303665",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7445255474452555",
      "ci_low": "0.6617619416187528",
      "ci_high": "0.8214285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.18421052631578946",
      "ci_low": "0.12631578947368421",
      "ci_high": "0.24848660010509718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.23157894736842105",
      "ci_low": "0.16756686005377106",
      "ci_high": "0.29907466075294353",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "68.21052631578948",
      "ci_low": "62.209944751381215",
      "ci_high": "74.03150546523376",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "45.05263157894737",
      "ci_low": "34.35776410209829",
      "ci_high": "55.44983015416775",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.3916102610839453",
      "ci_low": "0.31483962257213033",
      "ci_high": "0.41593097439525534",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.47757348912676256",
      "ci_low": "0.3839507592343053",
      "ci_high": "0.50723289560397",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.25842696629213485",
      "ci_low": "0.1574074074074074",
      "ci_high": "0.36082474226804123",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.42696629213483145",
      "ci_low": "0.3164524670627745",
      "ci_high": "0.5353535353535354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6052631578947368",
      "ci_low": "0.42857142857142855",
      "ci_high": "0.7619047619047619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.16853932584269662",
      "ci_low": "0.09410845588235295",
      "ci_high": "0.25",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2696629213483146",
      "ci_low": "0.17204301075268819",
      "ci_high": "0.375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "48.426966292134836",
      "ci_low": "40.602409638554214",
      "ci_high": "56.23535235876528",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "21.460674157303373",
      "ci_low": "6.321327350023619",
      "ci_high": "35.744813829787226",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.22526398332849945",
      "ci_low": "0.09734295722198946",
      "ci_high": "0.23848945409429279",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.31741743105379466",
      "ci_low": "0.1371650760855306",
      "ci_high": "0.3360533216783217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.125",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.125",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "9931",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "53.33333333333333",
      "ci_low": "27.5",
      "ci_high": "82.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "28.333333333333336",
      "ci_low": "-15.833333333333332",
      "ci_high": "75.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4629629629629629",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.5185185185185185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4629629629629629",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.5185185185185185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.24875621890547264",
      "ci_low": "0.1875",
      "ci_high": "0.31313270063270066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.417910447761194",
      "ci_low": "0.34972147063504694",
      "ci_high": "0.48444673539518895",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5952380952380952",
      "ci_low": "0.4861011904761905",
      "ci_high": "0.704225352112676",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.1691542288557214",
      "ci_low": "0.11737089201877934",
      "ci_high": "0.2215909090909091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.38308457711442784",
      "ci_low": "0.31527093596059114",
      "ci_high": "0.4489795918367347",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "50.09950248756219",
      "ci_low": "45.27358452303698",
      "ci_high": "55.000550660792946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "11.791044776119403",
      "ci_low": "1.7388764729784643",
      "ci_high": "21.980698590796766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.22384245871087974",
      "ci_low": "0.16039309283270098",
      "ci_high": "0.2486544112396175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3150375344819789",
      "ci_low": "0.22573842694972726",
      "ci_high": "0.3499580602631654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "-80.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.2",
      "ci_high": "0.4305555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.4857142857142857",
      "ci_low": "0.3580026455026455",
      "ci_high": "0.6067415730337079",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6470588235294118",
      "ci_low": "0.47364766081871346",
      "ci_high": "0.8064516129032258",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.17142857142857143",
      "ci_low": "0.0875",
      "ci_high": "0.2597402597402597",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.12162162162162163",
      "ci_high": "0.3150684931506849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "49.57142857142857",
      "ci_low": "40.0",
      "ci_high": "58.60018987341772",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "28.142857142857146",
      "ci_low": "13.035579004329005",
      "ci_high": "42.531846493871804",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2587624793507146",
      "ci_low": "0.16663461926861697",
      "ci_high": "0.3150431652898758",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.44359282174408227",
      "ci_low": "0.2856593473176291",
      "ci_high": "0.540073997639787",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.43902439024390244",
      "ci_high": "0.7575757575757576",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8444444444444444",
      "ci_low": "0.7297297297297297",
      "ci_high": "0.9444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7105263157894737",
      "ci_low": "0.534873696872494",
      "ci_high": "0.8780672579453065",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.24444444444444444",
      "ci_low": "0.10198979591836739",
      "ci_high": "0.40384615384615385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.26666666666666666",
      "ci_low": "0.11904761904761904",
      "ci_high": "0.42857142857142855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "72.22222222222223",
      "ci_low": "60.55518018018018",
      "ci_high": "83.27083333333331",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "45.55555555555556",
      "ci_low": "20.79923076923077",
      "ci_high": "68.9297161172161",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.3883116883116883",
      "ci_low": "0.2208509199134199",
      "ci_high": "0.40619295634920627",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5974025974025974",
      "ci_low": "0.33977064602064605",
      "ci_high": "0.6249122405372406",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.21333333333333335",
      "ci_low": "0.15289120863257424",
      "ci_high": "0.2777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.36",
      "ci_low": "0.29333173076923075",
      "ci_high": "0.4279284196860616",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5925925925925926",
      "ci_low": "0.47368421052631576",
      "ci_high": "0.7040816326530612",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.14666666666666667",
      "ci_low": "0.10344827586206896",
      "ci_high": "0.19262295081967212",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.36444444444444446",
      "ci_low": "0.2999884792626728",
      "ci_high": "0.42796653762990866",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "47.111111111111114",
      "ci_low": "42.29235275350502",
      "ci_high": "52.14971596115081",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "10.666666666666666",
      "ci_low": "0.8795901084010841",
      "ci_high": "20.627680988713475",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.1947866947866948",
      "ci_low": "0.14392780708570185",
      "ci_high": "0.22163805891279603",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.2813585591363369",
      "ci_low": "0.20789572134601372",
      "ci_high": "0.3201438628740388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.43661971830985913",
      "ci_low": "0.31942934782608695",
      "ci_high": "0.5573770491803278",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6197183098591549",
      "ci_low": "0.5",
      "ci_high": "0.7323943661971831",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7045454545454546",
      "ci_low": "0.5573315118397085",
      "ci_high": "0.8461864406779657",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.18309859154929578",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.2835820895522388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.29577464788732394",
      "ci_low": "0.18840579710144928",
      "ci_high": "0.410958904109589",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "61.40845070422535",
      "ci_low": "52.368238304093566",
      "ci_high": "70.17378526645767",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "31.83098591549296",
      "ci_low": "14.594483828090386",
      "ci_high": "48.571931589537215",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4093189964157706",
      "ci_low": "0.24628520225294417",
      "ci_high": "0.42804252199413484",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.46995884773662555",
      "ci_low": "0.28277189888300996",
      "ci_high": "0.49145622895622887",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6345514950166113",
      "ci_low": "0.5798045602605864",
      "ci_high": "0.6883128002134852",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7807308970099668",
      "ci_low": "0.7331285877734598",
      "ci_high": "0.8269230769230769",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8127659574468085",
      "ci_low": "0.7614641144712501",
      "ci_high": "0.861244019138756",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.1461794019933555",
      "ci_low": "0.10726643598615918",
      "ci_high": "0.1875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.19601328903654486",
      "ci_low": "0.15202702702702703",
      "ci_high": "0.24138224255219326",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "73.62126245847176",
      "ci_low": "69.4196423608409",
      "ci_high": "77.77787258248009",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "54.01993355481728",
      "ci_low": "46.4026061045595",
      "ci_high": "61.45289307164307",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.53739400667129",
      "ci_low": "0.46643933526127884",
      "ci_high": "0.5587819969490437",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6448728080055479",
      "ci_low": "0.5597272023135346",
      "ci_high": "0.6705383963388526",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6282051282051282",
      "ci_low": "0.5128205128205128",
      "ci_high": "0.7407407407407407",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8205128205128205",
      "ci_low": "0.726027397260274",
      "ci_high": "0.9069767441860465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.765625",
      "ci_low": "0.6461538461538462",
      "ci_high": "0.8727272727272727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.19230769230769232",
      "ci_low": "0.10446983830845771",
      "ci_high": "0.29411764705882354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.23076923076923078",
      "ci_low": "0.136986301369863",
      "ci_high": "0.33735316265060233",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "76.92307692307692",
      "ci_low": "69.11089743589744",
      "ci_high": "84.46153846153847",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "53.846153846153854",
      "ci_low": "37.05128205128205",
      "ci_high": "69.09120046620046",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4633477633477634",
      "ci_low": "0.3359846981721981",
      "ci_high": "0.5083335588023088",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6950216450216451",
      "ci_low": "0.5039770472582973",
      "ci_high": "0.7625003382034632",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5555555555555556",
      "ci_low": "0.4984423676012461",
      "ci_high": "0.6117656192911948",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7160493827160493",
      "ci_low": "0.6655390377031837",
      "ci_high": "0.7645262161297205",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7758620689655172",
      "ci_low": "0.7194547779586142",
      "ci_high": "0.8305785123966942",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.16049382716049382",
      "ci_low": "0.12084334926529047",
      "ci_high": "0.20327868852459016",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.21604938271604937",
      "ci_low": "0.1710914454277286",
      "ci_high": "0.2625080128205127",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "68.76543209876543",
      "ci_low": "64.44427083333333",
      "ci_high": "72.95693495143577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "47.160493827160494",
      "ci_low": "39.33923954165646",
      "ci_high": "54.75088696115828",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.3806293538084315",
      "ci_low": "0.3422577334462904",
      "ci_high": "0.3995077819102162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5709440307126472",
      "ci_low": "0.5133866001694357",
      "ci_high": "0.5992616728653243",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3026315789473684",
      "ci_low": "0.19696969696969696",
      "ci_high": "0.41333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.5394736842105263",
      "ci_low": "0.4230769230769231",
      "ci_high": "0.6538461538461539",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.5609756097560976",
      "ci_low": "0.4",
      "ci_high": "0.717948717948718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.33766233766233766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.42105263157894735",
      "ci_low": "0.3076923076923077",
      "ci_high": "0.5333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "50.65789473684211",
      "ci_low": "42.29881174277726",
      "ci_high": "59.25939542483659",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "8.552631578947368",
      "ci_low": "-8.588235294117647",
      "ci_high": "25.890586582367405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2020428861199652",
      "ci_low": "0.10126468333304835",
      "ci_high": "0.22335180145524974",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.29296218487394954",
      "ci_low": "0.1468337908329201",
      "ci_high": "0.3238601121101121",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5333333333333333",
      "ci_low": "0.42341476091476093",
      "ci_high": "0.6403508771929824",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6476190476190476",
      "ci_low": "0.5392156862745098",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8235294117647058",
      "ci_low": "0.7236842105263158",
      "ci_high": "0.9102564102564102",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.11428571428571428",
      "ci_low": "0.05714285714285714",
      "ci_high": "0.18085106382978725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.22857142857142856",
      "ci_low": "0.14423076923076922",
      "ci_high": "0.3191489361702128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "70.0",
      "ci_low": "62.64362394274561",
      "ci_high": "77.14285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "47.142857142857146",
      "ci_low": "32.47616690240452",
      "ci_high": "60.83355263157894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4924320898177107",
      "ci_low": "0.345452240660574",
      "ci_high": "0.5191738970905637",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5539861010449245",
      "ci_low": "0.38863377074314576",
      "ci_high": "0.5840706342268842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3824561403508772",
      "ci_low": "0.3231292517006803",
      "ci_high": "0.4416961130742049",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.543859649122807",
      "ci_low": "0.48288940830251575",
      "ci_high": "0.6043997827267789",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.7032258064516129",
      "ci_low": "0.625",
      "ci_high": "0.7777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.16140350877192983",
      "ci_low": "0.11787038294405214",
      "ci_high": "0.2080568099598313",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.28421052631578947",
      "ci_low": "0.2318840579710145",
      "ci_high": "0.3381308044473512",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "57.40350877192982",
      "ci_low": "52.75726672831684",
      "ci_high": "61.94144371529531",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "28.982456140350877",
      "ci_low": "20.545196108239587",
      "ci_high": "37.29454187588358",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.351231095671801",
      "ci_low": "0.2783211840908452",
      "ci_high": "0.3743689624053126",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.41427257438212434",
      "ci_low": "0.3282762684148431",
      "ci_high": "0.44156339155498414",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7061224489795919",
      "ci_low": "0.6459123702983139",
      "ci_high": "0.7647058823529411",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8408163265306122",
      "ci_low": "0.7890295358649789",
      "ci_high": "0.8870967741935484",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8398058252427184",
      "ci_low": "0.7887323943661971",
      "ci_high": "0.8883495145631068",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.1346938775510204",
      "ci_low": "0.0931174089068826",
      "ci_high": "0.17857142857142858",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.1836734693877551",
      "ci_low": "0.13692946058091288",
      "ci_high": "0.23423423423423423",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "82.9795918367347",
      "ci_low": "79.30128758064227",
      "ci_high": "86.52894874799556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "64.61224489795919",
      "ci_low": "56.44066101694915",
      "ci_high": "72.40528153644696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.5688163971718239",
      "ci_low": "0.4616772500720305",
      "ci_high": "0.5615022397967091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6845078677830421",
      "ci_low": "0.5555777077137993",
      "ci_high": "0.6757060851790907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5230769230769231",
      "ci_low": "0.4351145038167939",
      "ci_high": "0.6120689655172413",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7153846153846154",
      "ci_low": "0.6349206349206349",
      "ci_high": "0.7948717948717948",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7311827956989247",
      "ci_low": "0.6363636363636364",
      "ci_high": "0.8192771084337349",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.19230769230769232",
      "ci_low": "0.12666455696202533",
      "ci_high": "0.2631578947368421",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.3153846153846154",
      "ci_low": "0.23529411764705882",
      "ci_high": "0.3969465648854962",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "70.92307692307692",
      "ci_low": "65.1694209039548",
      "ci_high": "76.77426019728073",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "39.38461538461539",
      "ci_low": "26.315587044534414",
      "ci_high": "52.52104625634037",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4262057387057387",
      "ci_low": "0.3120593295593296",
      "ci_high": "0.43502816627816626",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5036976911976911",
      "ci_low": "0.36879738947920765",
      "ci_high": "0.5141241965105601",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.11842105263157894",
      "ci_low": "0.05061708860759495",
      "ci_high": "0.19444444444444445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14492276887871855",
      "ci_high": "0.33766233766233766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2631578947368421",
      "ci_high": "0.7333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.11842105263157894",
      "ci_low": "0.05128205128205128",
      "ci_high": "0.19753086419753085",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.15714285714285714",
      "ci_high": "0.35366322701688546",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "27.36842105263158",
      "ci_low": "20.416523972602736",
      "ci_high": "34.667088607594934",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "2.3684210526315788",
      "ci_low": "-11.14327122153209",
      "ci_high": "15.2",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.13647798742138365",
      "ci_low": "0.0440251572327044",
      "ci_high": "0.13836477987421383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.15724637681159417",
      "ci_low": "0.05072463768115942",
      "ci_high": "0.15942028985507245",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.20408163265306123",
      "ci_low": "0.12675610625779996",
      "ci_high": "0.2882882882882883",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.3469387755102041",
      "ci_low": "0.25263157894736843",
      "ci_high": "0.44554455445544555",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.5882352941176471",
      "ci_low": "0.42105263157894735",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.07692307692307693",
      "ci_high": "0.2127708825847123",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2857142857142857",
      "ci_low": "0.19801980198019803",
      "ci_high": "0.375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "34.38775510204082",
      "ci_low": "26.98108294025157",
      "ci_high": "42.17402066293586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "5.816326530612245",
      "ci_low": "-7.209358228980322",
      "ci_high": "18.991055045871555",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.1511175898931001",
      "ci_low": "0.08830911152339725",
      "ci_high": "0.17303759166004062",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.16097308488612835",
      "ci_low": "0.09406840140535792",
      "ci_high": "0.18432265198569542",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.4953271028037383",
      "ci_low": "0.40186915887850466",
      "ci_high": "0.5887850467289719",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7009345794392523",
      "ci_low": "0.6105263157894737",
      "ci_high": "0.7857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7066666666666667",
      "ci_low": "0.6025641025641025",
      "ci_high": "0.8076923076923077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.205607476635514",
      "ci_low": "0.13043478260869565",
      "ci_high": "0.2833333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2803738317757009",
      "ci_low": "0.19801980198019803",
      "ci_high": "0.36666666666666664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "66.54205607476635",
      "ci_low": "59.54015006385696",
      "ci_high": "73.16831683168317",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "38.50467289719626",
      "ci_low": "24.946369395711503",
      "ci_high": "51.681436605739805",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.46988304093567246",
      "ci_low": "0.2997446267183109",
      "ci_high": "0.4413140490114174",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5251633986928105",
      "ci_low": "0.335008700449877",
      "ci_high": "0.4932333488951136",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3308270676691729",
      "ci_low": "0.25",
      "ci_high": "0.4146341463414634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.518796992481203",
      "ci_low": "0.4297520661157025",
      "ci_high": "0.6083347902097901",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6376811594202898",
      "ci_low": "0.5147058823529411",
      "ci_high": "0.7534246575342466",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.18796992481203006",
      "ci_low": "0.12194297956493079",
      "ci_high": "0.2597402597402597",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3308270676691729",
      "ci_low": "0.2517482517482518",
      "ci_high": "0.4132231404958678",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "57.29323308270676",
      "ci_low": "51.267605633802816",
      "ci_high": "63.228034251675346",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "24.210526315789473",
      "ci_low": "11.904386951631047",
      "ci_high": "36.268686330253495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3162954217499672",
      "ci_low": "0.20259188523894406",
      "ci_high": "0.3204784244276222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3550254733928203",
      "ci_low": "0.22739905486003925",
      "ci_high": "0.3597206804799842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite LOW",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6013986013986014",
      "ci_low": "0.5514012835827887",
      "ci_high": "0.6498830935251798",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7808857808857809",
      "ci_low": "0.7410698227132579",
      "ci_high": "0.8198198198198198",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7701492537313432",
      "ci_low": "0.7226858731297396",
      "ci_high": "0.8164585049809003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.1794871794871795",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.21695819413729825",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.24708624708624707",
      "ci_low": "0.205377349019052",
      "ci_high": "0.2885106086214535",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "72.21445221445221",
      "ci_low": "68.62232779097387",
      "ci_high": "75.7792812026683",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "47.505827505827504",
      "ci_low": "40.381822865777146",
      "ci_high": "54.57332563510392",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4654919191330071",
      "ci_low": "0.3976692922665775",
      "ci_high": "0.4713525415265455",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6306664710834289",
      "ci_low": "0.5387777508127825",
      "ci_high": "0.6386066691649969",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.39058171745152354",
      "ci_low": "0.338623297087864",
      "ci_high": "0.44385026737967914",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5817174515235457",
      "ci_low": "0.5283018867924528",
      "ci_high": "0.6332475813544415",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.6714285714285714",
      "ci_low": "0.6048780487804878",
      "ci_high": "0.7355795704845816",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.19113573407202217",
      "ci_low": "0.15056818181818182",
      "ci_high": "0.23351648351648352",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.29085872576177285",
      "ci_low": "0.24331550802139038",
      "ci_high": "0.3408450704225352",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "59.11357340720222",
      "ci_low": "55.05679437531742",
      "ci_high": "63.083465189873415",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "30.027700831024934",
      "ci_low": "22.37971262401642",
      "ci_high": "37.378951486697964",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.2888468366428893",
      "ci_low": "0.22631365892797076",
      "ci_high": "0.29702985247806746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.39165672765137527",
      "ci_low": "0.3068659782074179",
      "ci_high": "0.4027523423431422",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6826666666666666",
      "ci_low": "0.6322713627765442",
      "ci_high": "0.7336818168918152",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.856",
      "ci_low": "0.8179271708683473",
      "ci_high": "0.8922662558436039",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7975077881619937",
      "ci_low": "0.7507148093841642",
      "ci_high": "0.8433734939759037",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.17333333333333334",
      "ci_low": "0.13350785340314136",
      "ci_high": "0.21448467966573817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.23466666666666666",
      "ci_low": "0.18994413407821228",
      "ci_high": "0.27979274611398963",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "81.81333333333333",
      "ci_low": "78.72109310580969",
      "ci_high": "84.86702127659575",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "58.346666666666664",
      "ci_low": "51.181061982636784",
      "ci_high": "65.44716486806813",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5561552774710669",
      "ci_low": "0.49417766617328757",
      "ci_high": "0.5620633601445499",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6773686071762993",
      "ci_low": "0.6018830549546451",
      "ci_high": "0.6845643488940031",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.20114942528735633",
      "ci_low": "0.143646408839779",
      "ci_high": "0.2648648648648649",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.41379310344827586",
      "ci_low": "0.3413173652694611",
      "ci_high": "0.4913294797687861",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.4861111111111111",
      "ci_low": "0.3723267866884888",
      "ci_high": "0.6027397260273972",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.21264367816091953",
      "ci_low": "0.15300546448087432",
      "ci_high": "0.2752808988764045",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.26282051282051283",
      "ci_high": "0.4037274370862858",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "34.88505747126437",
      "ci_low": "29.333240223463687",
      "ci_high": "40.574795694638226",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "1.5517241379310345",
      "ci_low": "-8.685714285714285",
      "ci_high": "11.850343208092475",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1834154351395731",
      "ci_low": "0.10519978106185005",
      "ci_high": "0.18124816514902722",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.20723562152133582",
      "ci_low": "0.1188620902906617",
      "ci_high": "0.2047868878956541",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.44583333333333336",
      "ci_low": "0.38235294117647056",
      "ci_high": "0.5084033613445378",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6291666666666667",
      "ci_low": "0.5670961850649351",
      "ci_high": "0.6901978136387298",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.7086092715231788",
      "ci_low": "0.6350259685569904",
      "ci_high": "0.781021897810219",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.18333333333333332",
      "ci_low": "0.13565891472868216",
      "ci_high": "0.23236742010628228",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2708333333333333",
      "ci_low": "0.21338772663877267",
      "ci_high": "0.3292181069958848",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "64.45833333333334",
      "ci_low": "59.95930653162464",
      "ci_high": "68.961186540732",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "37.375",
      "ci_low": "28.265651105256392",
      "ci_high": "46.22964093714741",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.428970606687998",
      "ci_low": "0.31851245222169133",
      "ci_high": "0.41735012252595366",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.49956070652273177",
      "ci_low": "0.3709258937265267",
      "ci_high": "0.4860279907897182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.55",
      "ci_low": "0.4987593052109181",
      "ci_high": "0.601467446561282",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7675",
      "ci_low": "0.7243086776218898",
      "ci_high": "0.8108761853987947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7166123778501629",
      "ci_low": "0.6643829691425321",
      "ci_high": "0.7682119205298014",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2175",
      "ci_low": "0.17585301837270342",
      "ci_high": "0.2595641551071879",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.28",
      "ci_low": "0.2347412100967664",
      "ci_high": "0.32673267326732675",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "68.925",
      "ci_low": "65.10467156042533",
      "ci_high": "72.69924807197943",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "40.925000000000004",
      "ci_low": "33.28179419525066",
      "ci_high": "48.51681318808937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.37299546410126566",
      "ci_low": "0.3405976362492402",
      "ci_high": "0.3908888539057059",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5967927425620252",
      "ci_low": "0.5449562179987842",
      "ci_high": "0.6254221662491294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.45897435897435895",
      "ci_low": "0.4067335979221605",
      "ci_high": "0.5117508891675746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.6102564102564103",
      "ci_low": "0.5604113110539846",
      "ci_high": "0.6595758715544826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.7521008403361344",
      "ci_low": "0.6945577638307764",
      "ci_high": "0.807088821016207",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.15128205128205127",
      "ci_low": "0.11622259668869839",
      "ci_high": "0.18932038834951456",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.25384615384615383",
      "ci_low": "0.20971763880867916",
      "ci_high": "0.3007528716249727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "63.46153846153846",
      "ci_low": "59.51079212660509",
      "ci_high": "67.33509234828496",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "38.07692307692307",
      "ci_low": "30.69584910849045",
      "ci_high": "45.163812138597315",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4142105923113062",
      "ci_low": "0.3694420747216249",
      "ci_high": "0.4390089536226359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5204184364936925",
      "ci_low": "0.46417081182973385",
      "ci_high": "0.5515753519874144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6611295681063123",
      "ci_low": "0.6075949367088608",
      "ci_high": "0.7147546617091994",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8305647840531561",
      "ci_low": "0.7875816993464052",
      "ci_high": "0.8719723183391004",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.796",
      "ci_low": "0.7447676488916585",
      "ci_high": "0.8454935622317596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16943521594684385",
      "ci_low": "0.1277258566978193",
      "ci_high": "0.21403538109487766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2159468438538206",
      "ci_low": "0.16943521594684385",
      "ci_high": "0.2647058823529412",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "75.74750830564784",
      "ci_low": "71.72630167877725",
      "ci_high": "79.75700708762886",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "54.152823920265774",
      "ci_low": "46.16599264705882",
      "ci_high": "62.075743179041666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5506170956666313",
      "ci_low": "0.467650702261773",
      "ci_high": "0.5677433357205305",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.688271369583289",
      "ci_low": "0.5845633778272165",
      "ci_high": "0.7096791696506631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2777777777777778",
      "ci_low": "0.21978021978021978",
      "ci_high": "0.33688038587661095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.4777777777777778",
      "ci_low": "0.41393405092175584",
      "ci_high": "0.5407436230004323",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.5813953488372093",
      "ci_low": "0.48717948717948717",
      "ci_high": "0.6742468287526426",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.1494661921708185",
      "ci_high": "0.25196850393700787",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3592592592592593",
      "ci_low": "0.29770992366412213",
      "ci_high": "0.4224834582076481",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "51.96296296296296",
      "ci_low": "47.45169899552369",
      "ci_high": "56.50791538630376",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "16.037037037037038",
      "ci_low": "7.096713329275716",
      "ci_high": "25.08491649593344",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.21416156230020156",
      "ci_low": "0.1778105954798358",
      "ci_high": "0.23847556640257622",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3252082983077135",
      "ci_low": "0.27000868202493594",
      "ci_high": "0.3621295637965045",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5707762557077626",
      "ci_low": "0.5",
      "ci_high": "0.6383954770288396",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7579908675799086",
      "ci_low": "0.6995884773662552",
      "ci_high": "0.8128085403440373",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7530120481927711",
      "ci_low": "0.6821182482869502",
      "ci_high": "0.8192790516906334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1872146118721461",
      "ci_low": "0.13592233009708737",
      "ci_high": "0.2415489130434782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.2237442922374429",
      "ci_low": "0.16901408450704225",
      "ci_high": "0.2809945919183455",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "70.73059360730593",
      "ci_low": "65.49994075829385",
      "ci_high": "75.76427338275616",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "48.35616438356165",
      "ci_low": "38.512740384615384",
      "ci_high": "57.69581406105457",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4859780640599667",
      "ci_low": "0.4031096651799534",
      "ci_high": "0.5110664250287895",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6479707520799556",
      "ci_low": "0.5374795535732712",
      "ci_high": "0.6814219000383861",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.660958904109589",
      "ci_low": "0.6062717770034843",
      "ci_high": "0.7153284671532847",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8356164383561644",
      "ci_low": "0.791970802919708",
      "ci_high": "0.8771929824561403",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7909836065573771",
      "ci_low": "0.7380926388022458",
      "ci_high": "0.8412698412698413",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.17465753424657535",
      "ci_low": "0.13157693545791083",
      "ci_high": "0.21993127147766323",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.2226027397260274",
      "ci_low": "0.17491749174917492",
      "ci_high": "0.2724017426770486",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "75.68493150684931",
      "ci_low": "71.53564895470383",
      "ci_high": "79.79464081183853",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "53.42465753424658",
      "ci_low": "45.18640929064658",
      "ci_high": "61.49714663643234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.5560774215026628",
      "ci_low": "0.4632518386966575",
      "ci_high": "0.5694879914973369",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6917060608935561",
      "ci_low": "0.5762400920373056",
      "ci_high": "0.7083875016186385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5947368421052631",
      "ci_low": "0.5174411373365729",
      "ci_high": "0.6687509952229299",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7631578947368421",
      "ci_low": "0.6972960575254153",
      "ci_high": "0.826530612244898",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7793103448275862",
      "ci_low": "0.7101449275362319",
      "ci_high": "0.8435374149659864",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.16842105263157894",
      "ci_low": "0.11827956989247312",
      "ci_high": "0.2227986314485244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.14507577720207254",
      "ci_high": "0.25842696629213485",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "72.42105263157895",
      "ci_low": "66.50536329470717",
      "ci_high": "77.99045610706594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "52.42105263157895",
      "ci_low": "42.471670148495285",
      "ci_high": "61.73924191750279",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4310133470826493",
      "ci_low": "0.34643927565894855",
      "ci_high": "0.44500373202994803",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.546651074348726",
      "ci_low": "0.4393863983967151",
      "ci_high": "0.5643949772087146",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2696629213483146",
      "ci_low": "0.16666666666666666",
      "ci_high": "0.375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.5056179775280899",
      "ci_low": "0.39361702127659576",
      "ci_high": "0.6133333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.5333333333333333",
      "ci_low": "0.36363636363636365",
      "ci_high": "0.6904761904761905",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.23595505617977527",
      "ci_low": "0.14634146341463414",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.34831460674157305",
      "ci_low": "0.24358695652173912",
      "ci_high": "0.45977011494252873",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "51.79775280898877",
      "ci_low": "43.732696629213486",
      "ci_high": "59.791824494949495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "16.96629213483146",
      "ci_low": "0.2324913124832933",
      "ci_high": "33.09532193158954",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.20684534320897957",
      "ci_low": "0.09965570855716507",
      "ci_high": "0.22044746353046749",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3102680148134694",
      "ci_low": "0.1494835628357476",
      "ci_high": "0.3306711952957012",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.07692307692307693",
      "ci_high": "0.6666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.75",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "9817",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.08333333333333333",
      "ci_high": "0.6",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "49.16666666666667",
      "ci_low": "30.554824561403507",
      "ci_high": "71.42857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "15.833333333333332",
      "ci_low": "-21.11458333333333",
      "ci_high": "56.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3",
      "ci_low": "0.0",
      "ci_high": "0.3",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.30845771144278605",
      "ci_low": "0.245",
      "ci_high": "0.375014140271493",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5124378109452736",
      "ci_low": "0.4427083333333333",
      "ci_high": "0.5816351059654631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6019417475728155",
      "ci_low": "0.5",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.20398009950248755",
      "ci_low": "0.1477832512315271",
      "ci_high": "0.2644230769230769",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.36318407960199006",
      "ci_low": "0.2943856514568444",
      "ci_high": "0.43315508021390375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "53.830845771144276",
      "ci_low": "48.94472361809045",
      "ci_high": "58.895151933701655",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "17.512437810945272",
      "ci_low": "7.178217821782178",
      "ci_high": "27.819616158711895",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.24490406888433203",
      "ci_low": "0.18414353873358325",
      "ci_high": "0.26831842938557754",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.362820842791603",
      "ci_low": "0.27280524256827143",
      "ci_high": "0.3975087842749297",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6336",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.19148153942428037",
      "ci_high": "0.4366417253521123",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.5714285714285714",
      "ci_low": "0.44285714285714284",
      "ci_high": "0.6901408450704225",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.55",
      "ci_low": "0.3725376196990424",
      "ci_high": "0.717948717948718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2571428571428571",
      "ci_low": "0.15254237288135594",
      "ci_high": "0.36923076923076925",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.20408163265306123",
      "ci_high": "0.4264705882352941",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "51.42857142857143",
      "ci_low": "41.738755622188904",
      "ci_high": "61.03896103896104",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "20.0",
      "ci_low": "1.9991071428571436",
      "ci_high": "37.4656690140845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.20805120805120805",
      "ci_low": "0.13358375420875418",
      "ci_high": "0.2565779570733131",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.4458240172525887",
      "ci_low": "0.28625090187590185",
      "ci_high": "0.5498099080142423",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7111111111111111",
      "ci_low": "0.5686274509803921",
      "ci_high": "0.8484991423670667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8666666666666667",
      "ci_low": "0.7619047619047619",
      "ci_high": "0.9583333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8205128205128205",
      "ci_low": "0.6888888888888889",
      "ci_high": "0.9354838709677419",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.15555555555555556",
      "ci_low": "0.05660377358490566",
      "ci_high": "0.26666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.15555555555555556",
      "ci_low": "0.05660377358490566",
      "ci_high": "0.26666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "79.11111111111111",
      "ci_low": "67.85666023166024",
      "ci_high": "89.42857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "63.55555555555555",
      "ci_low": "43.80667701863354",
      "ci_high": "81.9447463768116",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4874458874458875",
      "ci_low": "0.2865454076367389",
      "ci_high": "0.5018803802985726",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7124209124209125",
      "ci_low": "0.4187971342383107",
      "ci_high": "0.7335174788979137",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.24888888888888888",
      "ci_low": "0.18686868686868688",
      "ci_high": "0.3150684931506849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.4577777777777778",
      "ci_low": "0.38766256349724937",
      "ci_high": "0.5270306991774382",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5436893203883495",
      "ci_low": "0.4361633933442444",
      "ci_high": "0.6516879012841091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.2088888888888889",
      "ci_low": "0.15195923306031273",
      "ci_high": "0.26724237247735577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.37777777777777777",
      "ci_low": "0.3090909090909091",
      "ci_high": "0.44782608695652176",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "50.666666666666664",
      "ci_low": "45.84345846600748",
      "ci_high": "55.61090444082771",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "12.88888888888889",
      "ci_low": "2.991228070175439",
      "ci_high": "22.827031026740137",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.2171487138592402",
      "ci_low": "0.17080374912611754",
      "ci_high": "0.2437081310502363",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3056167083944861",
      "ci_low": "0.24039046173305437",
      "ci_high": "0.3429966288855178",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5492957746478874",
      "ci_low": "0.43333333333333335",
      "ci_high": "0.6625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6619718309859155",
      "ci_low": "0.5492888704777686",
      "ci_high": "0.7704918032786885",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8297872340425532",
      "ci_low": "0.7115384615384616",
      "ci_high": "0.9318181818181818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.11267605633802817",
      "ci_low": "0.04411764705882353",
      "ci_high": "0.1927710843373494",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.2112676056338028",
      "ci_low": "0.12",
      "ci_high": "0.3103448275862069",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "68.16901408450704",
      "ci_low": "59.20634920634921",
      "ci_high": "76.86567164179104",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "47.04225352112676",
      "ci_low": "30.75649350649351",
      "ci_high": "62.17405239687849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5015873015873016",
      "ci_low": "0.3269027777777778",
      "ci_high": "0.5134478021978022",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5573192239858906",
      "ci_low": "0.36322530864197533",
      "ci_high": "0.570497557997558",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6611295681063123",
      "ci_low": "0.6075949367088608",
      "ci_high": "0.7147546617091994",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8305647840531561",
      "ci_low": "0.7875816993464052",
      "ci_high": "0.8719723183391004",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.796",
      "ci_low": "0.7447676488916585",
      "ci_high": "0.8454935622317596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.16943521594684385",
      "ci_low": "0.1277258566978193",
      "ci_high": "0.21403538109487766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2159468438538206",
      "ci_low": "0.16943521594684385",
      "ci_high": "0.2647058823529412",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "75.74750830564784",
      "ci_low": "71.72630167877725",
      "ci_high": "79.75700708762886",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "54.152823920265774",
      "ci_low": "46.16599264705882",
      "ci_high": "62.075743179041666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5506170956666313",
      "ci_low": "0.467650702261773",
      "ci_high": "0.5677433357205305",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.688271369583289",
      "ci_low": "0.5845633778272165",
      "ci_high": "0.7096791696506631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6538461538461539",
      "ci_low": "0.539304235090752",
      "ci_high": "0.7619047619047619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8461538461538461",
      "ci_low": "0.7638706140350877",
      "ci_high": "0.9206349206349206",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7727272727272727",
      "ci_low": "0.6617590497737557",
      "ci_high": "0.8703703703703703",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.19230769230769232",
      "ci_low": "0.10989010989010989",
      "ci_high": "0.2876764217517641",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.21794871794871795",
      "ci_low": "0.13157894736842105",
      "ci_high": "0.3181818181818182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "78.46153846153845",
      "ci_low": "70.67790884104444",
      "ci_high": "85.76574952561668",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "56.66666666666667",
      "ci_low": "40.847104519774014",
      "ci_high": "71.01307706626953",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5636313004734058",
      "ci_low": "0.3960316657027183",
      "ci_high": "0.6053564579739392",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.739766081871345",
      "ci_low": "0.5197915612348178",
      "ci_high": "0.7945303510907952",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5987654320987654",
      "ci_low": "0.541795665634675",
      "ci_high": "0.6547619047619048",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7901234567901234",
      "ci_low": "0.7460802607354332",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7578125",
      "ci_low": "0.7017543859649122",
      "ci_high": "0.8125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.19135802469135801",
      "ci_low": "0.14733475716429706",
      "ci_high": "0.2367601246105919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.24382716049382716",
      "ci_low": "0.1962025316455696",
      "ci_high": "0.2920661216293746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "71.9753086419753",
      "ci_low": "67.77447845045957",
      "ci_high": "76.1583729077462",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "47.592592592592595",
      "ci_low": "39.2992409440819",
      "ci_high": "55.722981670270826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.414646816983449",
      "ci_low": "0.372733505192438",
      "ci_high": "0.4361010420009951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.6110584671335039",
      "ci_low": "0.5492914813362246",
      "ci_high": "0.6426752197909402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.34210526315789475",
      "ci_low": "0.22972972972972974",
      "ci_high": "0.4597837985769018",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6710526315789473",
      "ci_low": "0.548376049491825",
      "ci_high": "0.7875094696969696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.5098039215686274",
      "ci_low": "0.36363636363636365",
      "ci_high": "0.6491228070175439",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.32894736842105265",
      "ci_low": "0.22093023255813954",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.4342105263157895",
      "ci_low": "0.3176470588235294",
      "ci_high": "0.5584415584415584",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "55.921052631578945",
      "ci_low": "47.42825040128411",
      "ci_high": "64.45792400370715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "12.5",
      "ci_low": "-5.946687541199736",
      "ci_high": "30.735430283224403",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.19913419913419914",
      "ci_low": "0.1034880018970928",
      "ci_high": "0.21906589080573632",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3285714285714286",
      "ci_low": "0.17075520313020315",
      "ci_high": "0.361458719829465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6095238095238096",
      "ci_low": "0.5045871559633027",
      "ci_high": "0.7118838885641922",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.7523809523809524",
      "ci_low": "0.6630404250122129",
      "ci_high": "0.8372093023255814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.810126582278481",
      "ci_low": "0.7162162162162162",
      "ci_high": "0.8939438502673795",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.08",
      "ci_high": "0.21052631578947367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.2571428571428571",
      "ci_low": "0.17391304347826086",
      "ci_high": "0.34375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "72.95238095238095",
      "ci_low": "65.64697146185206",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "47.23809523809524",
      "ci_low": "32.27272727272727",
      "ci_high": "61.66745283018867",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5671821234321234",
      "ci_low": "0.401583014438673",
      "ci_high": "0.5820668544807199",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6617124773374773",
      "ci_low": "0.4685135168451185",
      "ci_high": "0.6790779968941733",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.40350877192982454",
      "ci_low": "0.3444776238254499",
      "ci_high": "0.462688030436055",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5578947368421052",
      "ci_low": "0.5",
      "ci_high": "0.6167280761477565",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.7232704402515723",
      "ci_low": "0.6511588887154675",
      "ci_high": "0.7922077922077922",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.1543859649122807",
      "ci_low": "0.11347517730496454",
      "ci_high": "0.19798657718120805",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.25263157894736843",
      "ci_low": "0.20149185391722704",
      "ci_high": "0.3055555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "59.96491228070176",
      "ci_low": "55.41241290795747",
      "ci_high": "64.43727005150846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "34.70175438596491",
      "ci_low": "26.573350836888743",
      "ci_high": "42.42138157894737",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3451476581410792",
      "ci_low": "0.2822105819473091",
      "ci_high": "0.36753220941886483",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4247971177120975",
      "ci_low": "0.3473361008582265",
      "ci_high": "0.45234733466937216",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7428571428571429",
      "ci_low": "0.6818139257745999",
      "ci_high": "0.803347280334728",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9102040816326531",
      "ci_low": "0.8710905997983871",
      "ci_high": "0.9444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8161434977578476",
      "ci_low": "0.7592592592592593",
      "ci_high": "0.870590770791075",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.1673469387755102",
      "ci_low": "0.11764705882352941",
      "ci_high": "0.21900924832743013",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.21224489795918366",
      "ci_low": "0.1582966207966208",
      "ci_high": "0.2681992337164751",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "85.10204081632652",
      "ci_low": "81.29028673835124",
      "ci_high": "88.7148778600641",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "63.87755102040816",
      "ci_low": "54.79495281719245",
      "ci_high": "72.6584388185654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.605582264957265",
      "ci_low": "0.4907888170970279",
      "ci_high": "0.593548859619793",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7390156453715775",
      "ci_low": "0.5989287259489153",
      "ci_high": "0.7243308117394084",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5692307692307692",
      "ci_low": "0.48333333333333334",
      "ci_high": "0.6571526418786692",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7538461538461538",
      "ci_low": "0.6744186046511628",
      "ci_high": "0.8321180644173345",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7551020408163265",
      "ci_low": "0.6666666666666666",
      "ci_high": "0.8387096774193549",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.18461538461538463",
      "ci_low": "0.12",
      "ci_high": "0.2540983606557377",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.27692307692307694",
      "ci_low": "0.2",
      "ci_high": "0.3577981651376147",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "75.61538461538461",
      "ci_low": "70.18158307210032",
      "ci_high": "81.01578182025547",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "47.92307692307692",
      "ci_low": "35.65859832305015",
      "ci_high": "59.60629921259843",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.45083943833943835",
      "ci_low": "0.34779137529137527",
      "ci_high": "0.45883793290043284",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5328102453102453",
      "ci_low": "0.411026170798898",
      "ci_high": "0.5422630116096026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.18421052631578946",
      "ci_low": "0.10168877425693934",
      "ci_high": "0.2753623188405797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.40789473684210525",
      "ci_low": "0.2987012987012987",
      "ci_high": "0.5185185185185185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.45161290322580644",
      "ci_low": "0.27586206896551724",
      "ci_high": "0.6296296296296297",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.2236842105263158",
      "ci_low": "0.13333333333333333",
      "ci_high": "0.32",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.32894736842105265",
      "ci_low": "0.2261904761904762",
      "ci_high": "0.43478260869565216",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "33.026315789473685",
      "ci_low": "25.0",
      "ci_high": "41.28246753246753",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.13157894736842105",
      "ci_low": "-15.238721804511277",
      "ci_high": "15.600624999999992",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.1812121212121212",
      "ci_low": "0.07454329004329004",
      "ci_high": "0.1777215007215007",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.21666666666666667",
      "ci_low": "0.08912784679089027",
      "ci_high": "0.21249309868875085",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.13592233009708737",
      "ci_high": "0.3020893895348836",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.41836734693877553",
      "ci_low": "0.31868131868131866",
      "ci_high": "0.5252525252525253",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.5121951219512195",
      "ci_low": "0.358974358974359",
      "ci_high": "0.6666666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.20408163265306123",
      "ci_low": "0.125",
      "ci_high": "0.28846153846153844",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.336734693877551",
      "ci_low": "0.24443798449612403",
      "ci_high": "0.4326923076923077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "36.326530612244895",
      "ci_low": "28.942307692307693",
      "ci_high": "44.41860465116279",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "2.653061224489796",
      "ci_low": "-11.087179868969624",
      "ci_high": "16.551819923371646",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.15296495956873316",
      "ci_low": "0.08513426039369436",
      "ci_high": "0.17103250915750914",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.17624223602484473",
      "ci_low": "0.09808947393186523",
      "ci_high": "0.19705919533365185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5700934579439252",
      "ci_low": "0.47572019735795",
      "ci_high": "0.66087324981577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7476635514018691",
      "ci_low": "0.6630434782608695",
      "ci_high": "0.8285714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7625",
      "ci_low": "0.6626506024096386",
      "ci_high": "0.8536653877774731",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.17757009345794392",
      "ci_low": "0.10678578806031812",
      "ci_high": "0.25555764411027565",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.27102803738317754",
      "ci_low": "0.18867924528301888",
      "ci_high": "0.35714285714285715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "70.2803738317757",
      "ci_low": "63.33333333333333",
      "ci_high": "76.96449049803407",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "43.177570093457945",
      "ci_low": "28.761607142857144",
      "ci_high": "57.07885074758913",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5385204081632653",
      "ci_low": "0.34609539270253564",
      "ci_high": "0.502640306122449",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5913165266106443",
      "ci_low": "0.38002631355572525",
      "ci_high": "0.5519187675070027",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3458646616541353",
      "ci_low": "0.2635556915544676",
      "ci_high": "0.42857142857142855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5338345864661654",
      "ci_low": "0.4485294117647059",
      "ci_high": "0.6166666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.647887323943662",
      "ci_low": "0.5342465753424658",
      "ci_high": "0.7567567567567568",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.18796992481203006",
      "ci_low": "0.125",
      "ci_high": "0.2565833926031294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.2706766917293233",
      "ci_low": "0.19548872180451127",
      "ci_high": "0.34965909090909075",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "59.774436090225564",
      "ci_low": "53.90243902439025",
      "ci_high": "65.54744525547446",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "32.70676691729324",
      "ci_low": "21.35961344537815",
      "ci_high": "43.68870277975766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.31372074253430193",
      "ci_low": "0.20401790300095382",
      "ci_high": "0.31341736467517023",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.37774538386783285",
      "ci_low": "0.2456542097358424",
      "ci_high": "0.3773800921598988",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite MEDIUM",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6060606060606061",
      "ci_low": "0.5576459745188398",
      "ci_high": "0.654320987654321",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8135198135198135",
      "ci_low": "0.7748799387214119",
      "ci_high": "0.8495151249550521",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7449856733524355",
      "ci_low": "0.6974063400576369",
      "ci_high": "0.7916666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.20745920745920746",
      "ci_low": "0.16904761904761906",
      "ci_high": "0.24708624708624707",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.2564102564102564",
      "ci_low": "0.21461852125483066",
      "ci_high": "0.29904306220095694",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "72.82051282051282",
      "ci_low": "69.22169094018271",
      "ci_high": "76.29631813417191",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "47.17948717948718",
      "ci_low": "39.95540736520433",
      "ci_high": "54.11058386258661",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4296899965310341",
      "ci_low": "0.37659909514081835",
      "ci_high": "0.44071373632432786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.602952091906451",
      "ci_low": "0.5284535689879226",
      "ci_high": "0.6184208880680085",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.40443213296398894",
      "ci_low": "0.352112676056338",
      "ci_high": "0.4568245125348189",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.6343490304709142",
      "ci_low": "0.5826233626609914",
      "ci_high": "0.685016980080085",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.6375545851528385",
      "ci_low": "0.5714285714285714",
      "ci_high": "0.7021276595744681",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.2299168975069252",
      "ci_low": "0.18579234972677597",
      "ci_high": "0.2755681818181818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3379501385041551",
      "ci_low": "0.28653216194242503",
      "ci_high": "0.38875035521454954",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "59.750692520775615",
      "ci_low": "55.69862248412031",
      "ci_high": "63.73276607072585",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "25.955678670360108",
      "ci_low": "17.95419130327667",
      "ci_high": "33.704918575400505",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.2868774052367834",
      "ci_low": "0.2262036091715442",
      "ci_high": "0.2938424611300952",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4084356277947424",
      "ci_low": "0.3220525961086392",
      "ci_high": "0.4183519785581017",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6906666666666667",
      "ci_low": "0.6416181349327075",
      "ci_high": "0.7393787036321076",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8746666666666667",
      "ci_low": "0.8404255319148937",
      "ci_high": "0.9076527298558961",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7896341463414634",
      "ci_low": "0.743202416918429",
      "ci_high": "0.833846394984326",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.184",
      "ci_low": "0.14477123606603642",
      "ci_high": "0.22493224932249323",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.23466666666666666",
      "ci_low": "0.19137466307277629",
      "ci_high": "0.27823691460055094",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "82.45333333333333",
      "ci_low": "79.51025646442454",
      "ci_high": "85.38482384823848",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "58.986666666666665",
      "ci_low": "52.010430906581846",
      "ci_high": "65.8529431063764",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5346079846079846",
      "ci_low": "0.4735553460474888",
      "ci_high": "0.5418435870400683",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6785409035409036",
      "ci_low": "0.6010510161371972",
      "ci_high": "0.687724552781625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1839080459770115",
      "ci_low": "0.12727272727272726",
      "ci_high": "0.2441893498942917",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.45977011494252873",
      "ci_low": "0.3864995817066369",
      "ci_high": "0.534893980848153",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.29333333333333333",
      "ci_high": "0.5068493150684932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.27586206896551724",
      "ci_low": "0.21176470588235294",
      "ci_high": "0.34355828220858897",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.39080459770114945",
      "ci_low": "0.3172043010752688",
      "ci_high": "0.46703296703296704",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "34.54022988505747",
      "ci_low": "29.161550393189525",
      "ci_high": "40.127471795512264",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "-4.540229885057471",
      "ci_low": "-14.904488643191923",
      "ci_high": "5.9395202020202005",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.16573100812231245",
      "ci_low": "0.09024741034931252",
      "ci_high": "0.16307320383687074",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.19801626944484085",
      "ci_low": "0.10782807470307469",
      "ci_high": "0.19484071107781958",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.475",
      "ci_low": "0.4126984126984127",
      "ci_high": "0.5381526104417671",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.7041666666666667",
      "ci_low": "0.643936799501868",
      "ci_high": "0.7615070807853235",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6745562130177515",
      "ci_low": "0.6031746031746031",
      "ci_high": "0.7430205757548005",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.22916666666666666",
      "ci_low": "0.17721394398485",
      "ci_high": "0.28270042194092826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.31666666666666665",
      "ci_low": "0.2562999166608792",
      "ci_high": "0.3782612766650382",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "65.75",
      "ci_low": "61.14406779661017",
      "ci_high": "70.3004615583839",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "34.08333333333333",
      "ci_low": "24.568741379310346",
      "ci_high": "43.468783225967854",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.452971341681019",
      "ci_low": "0.3382683254267695",
      "ci_high": "0.43860563197256747",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5332447440042375",
      "ci_low": "0.3982146109454374",
      "ci_high": "0.5163332123221365",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.56",
      "ci_low": "0.5099502487562189",
      "ci_high": "0.6092575112467866",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.8075",
      "ci_low": "0.7692307692307693",
      "ci_high": "0.8445040214477212",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6934984520123839",
      "ci_low": "0.6426424072975797",
      "ci_high": "0.7435897435897436",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2475",
      "ci_low": "0.206266318537859",
      "ci_high": "0.2889541525518862",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.305",
      "ci_low": "0.25925925925925924",
      "ci_high": "0.35189980951210514",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "69.45",
      "ci_low": "65.65725921183666",
      "ci_high": "73.16758911457865",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "38.95",
      "ci_low": "31.123242862629624",
      "ci_high": "46.561740359378156",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.36069291156813826",
      "ci_low": "0.32860588250455736",
      "ci_high": "0.3773509533405161",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6131779496658349",
      "ci_low": "0.5586300002577476",
      "ci_high": "0.6414966206788774",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4666666666666667",
      "ci_low": "0.4143920595533499",
      "ci_high": "0.5194830626292767",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.6538461538461539",
      "ci_low": "0.6036269430051814",
      "ci_high": "0.7029984854562304",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.7137254901960784",
      "ci_low": "0.6521739130434783",
      "ci_high": "0.7706785491396475",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.18717948717948718",
      "ci_low": "0.14761904761904762",
      "ci_high": "0.22928258427700063",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.28205128205128205",
      "ci_low": "0.23466340206185568",
      "ci_high": "0.3317079033771107",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "64.17948717948718",
      "ci_low": "60.223282101688866",
      "ci_high": "68.08632116131773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "35.97435897435898",
      "ci_low": "28.350197617534068",
      "ci_high": "43.359794206533586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.3922962036147562",
      "ci_low": "0.34947100437239237",
      "ci_high": "0.41544845062666175",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5230616048196749",
      "ci_low": "0.46596133916318977",
      "ci_high": "0.5539312675022157",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6578073089700996",
      "ci_low": "0.6028356438579583",
      "ci_high": "0.7105284312066339",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8438538205980066",
      "ci_low": "0.8027199508565969",
      "ci_high": "0.8836496270293989",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7795275590551181",
      "ci_low": "0.7265625",
      "ci_high": "0.8292707123499806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.18604651162790697",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.2317470829391359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.23588039867109634",
      "ci_low": "0.18855218855218855",
      "ci_high": "0.28619557645419713",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "75.28239202657807",
      "ci_low": "71.14580702861953",
      "ci_high": "79.35837467450838",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "51.69435215946844",
      "ci_low": "43.48546130634887",
      "ci_high": "59.72053872053872",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.548301740914744",
      "ci_low": "0.4666071433918518",
      "ci_high": "0.5655581090191384",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6853771761434301",
      "ci_low": "0.5832589292398148",
      "ci_high": "0.7069476362739231",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3037037037037037",
      "ci_low": "0.24302788844621515",
      "ci_high": "0.36462093862815886",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.5370370370370371",
      "ci_low": "0.47547169811320755",
      "ci_high": "0.5977454013261937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.5655172413793104",
      "ci_low": "0.47435897435897434",
      "ci_high": "0.6535947712418301",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.23333333333333334",
      "ci_low": "0.18076923076923077",
      "ci_high": "0.28825622775800713",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.37777777777777777",
      "ci_low": "0.3134905828002843",
      "ci_high": "0.44285837864557587",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "54.148148148148145",
      "ci_low": "49.4443407960199",
      "ci_high": "58.82575757575758",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "16.37037037037037",
      "ci_low": "6.6539015151515155",
      "ci_high": "26.103201655155985",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.20128487015549792",
      "ci_low": "0.16713878059185108",
      "ci_high": "0.22180235468413248",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3503847739743853",
      "ci_low": "0.29094528473396297",
      "ci_high": "0.3861003951908972",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5753424657534246",
      "ci_low": "0.5070422535211268",
      "ci_high": "0.6414165237278445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.817351598173516",
      "ci_low": "0.7646998055420515",
      "ci_high": "0.866078058698727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7039106145251397",
      "ci_low": "0.632768361581921",
      "ci_high": "0.7719298245614035",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.2420091324200913",
      "ci_low": "0.18551649928263989",
      "ci_high": "0.30180392656807753",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.2694063926940639",
      "ci_low": "0.21081014838367781",
      "ci_high": "0.3316345196615231",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "70.91324200913242",
      "ci_low": "65.75732795242142",
      "ci_high": "75.78205959237457",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "43.97260273972603",
      "ci_low": "33.807226385837964",
      "ci_high": "53.67736769233948",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4603288322631438",
      "ci_low": "0.3793467720466918",
      "ci_high": "0.4882993901812909",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6416704934577155",
      "ci_low": "0.528786409519631",
      "ci_high": "0.6806597560102842",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6541095890410958",
      "ci_low": "0.5984744957614733",
      "ci_high": "0.7075827266512347",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8424657534246576",
      "ci_low": "0.8006872852233677",
      "ci_high": "0.8827586206896552",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7764227642276422",
      "ci_low": "0.7219917012448133",
      "ci_high": "0.8275862068965517",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.18835616438356165",
      "ci_low": "0.1444024712493952",
      "ci_high": "0.23529411764705882",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.23972602739726026",
      "ci_low": "0.19093851132686085",
      "ci_high": "0.2905411779704232",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "75.1027397260274",
      "ci_low": "70.83605072463769",
      "ci_high": "79.31037309214246",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "51.13013698630137",
      "ci_low": "42.69230769230769",
      "ci_high": "59.389376047280464",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.552189819663868",
      "ci_low": "0.46020851912179145",
      "ci_high": "0.5664852154343418",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6868702634843237",
      "ci_low": "0.5724544993953992",
      "ci_high": "0.7046523411500352",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5894736842105263",
      "ci_low": "0.5116245656241647",
      "ci_high": "0.6630434782608695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8157894736842105",
      "ci_low": "0.7524752475247525",
      "ci_high": "0.8742864949258392",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7225806451612903",
      "ci_low": "0.6466666666666666",
      "ci_high": "0.7919463087248322",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.22631578947368422",
      "ci_low": "0.1685363690124187",
      "ci_high": "0.28846153846153844",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.2631578947368421",
      "ci_low": "0.19895287958115182",
      "ci_high": "0.3317079033771107",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "71.47368421052632",
      "ci_low": "65.442202731234",
      "ci_high": "77.10394035157022",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "45.15789473684211",
      "ci_low": "34.06375733855186",
      "ci_high": "55.63996065304204",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4332091777736967",
      "ci_low": "0.3324989172799802",
      "ci_high": "0.4450339755671601",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5600021078538031",
      "ci_low": "0.4298156735570476",
      "ci_high": "0.5752878220746216",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.34831460674157305",
      "ci_low": "0.24675324675324675",
      "ci_high": "0.45054945054945056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6292134831460674",
      "ci_low": "0.5154639175257731",
      "ci_high": "0.7333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.5535714285714286",
      "ci_low": "0.4218544407894737",
      "ci_high": "0.6818181818181818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2808988764044944",
      "ci_low": "0.18947368421052632",
      "ci_high": "0.3763440860215054",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.34831460674157305",
      "ci_low": "0.25",
      "ci_high": "0.45161290322580644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "55.95505617977528",
      "ci_low": "48.04597701149425",
      "ci_high": "63.626623376623364",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "21.123595505617978",
      "ci_low": "5.2069909793814455",
      "ci_high": "36.11764705882353",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.24125424125424125",
      "ci_low": "0.1280017445458622",
      "ci_high": "0.248893117018117",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.405745769382133",
      "ci_low": "0.21527566128167733",
      "ci_high": "0.4185929695304695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.21428571428571427",
      "ci_high": "0.7777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "9979",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5833333333333334",
      "ci_low": "0.25",
      "ci_high": "0.866874999999997",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "49.16666666666667",
      "ci_low": "30.554824561403507",
      "ci_high": "71.42857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "-9.166666666666666",
      "ci_low": "-50.0",
      "ci_high": "40.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.2727272727272727",
      "ci_low": "0.0",
      "ci_high": "0.2727272727272727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.32338308457711445",
      "ci_low": "0.2564039408866995",
      "ci_high": "0.38974358974358975",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5472636815920398",
      "ci_low": "0.47752456100728363",
      "ci_high": "0.6161137440758294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5909090909090909",
      "ci_low": "0.49019607843137253",
      "ci_high": "0.688",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.22388059701492538",
      "ci_low": "0.16304347826086957",
      "ci_high": "0.28643281511725294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.36318407960199006",
      "ci_low": "0.29523809523809524",
      "ci_high": "0.4336734693877551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "56.21890547263682",
      "ci_low": "51.179487179487175",
      "ci_high": "61.111258865248224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "19.900497512437813",
      "ci_low": "9.368919264179867",
      "ci_high": "30.154120879120875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.231350419918409",
      "ci_low": "0.17605690753175274",
      "ci_high": "0.25132825836997486",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.37701549912629617",
      "ci_low": "0.2869075530147082",
      "ci_high": "0.409571976602922",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "10.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "-90.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.37142857142857144",
      "ci_low": "0.2542372881355932",
      "ci_high": "0.4875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6857142857142857",
      "ci_low": "0.5692307692307692",
      "ci_high": "0.7903225806451613",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.5416666666666666",
      "ci_low": "0.39215686274509803",
      "ci_high": "0.6829268292682927",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.20754716981132076",
      "ci_high": "0.42857142857142855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.34285714285714286",
      "ci_low": "0.23376623376623376",
      "ci_high": "0.4576271186440678",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "45.713833634719705",
      "ci_high": "64.06263241525424",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "20.714285714285715",
      "ci_low": "2.638706140350877",
      "ci_high": "38.20908848614072",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.20425435131317482",
      "ci_low": "0.14027708198664082",
      "ci_high": "0.24297803343998994",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.5252254748053068",
      "ci_low": "0.3607124965370764",
      "ci_high": "0.624800657417117",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7333333333333333",
      "ci_low": "0.5909090909090909",
      "ci_high": "0.8648648648648649",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8888888888888888",
      "ci_low": "0.7916666666666666",
      "ci_high": "0.9743589743589743",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.825",
      "ci_low": "0.7",
      "ci_high": "0.9375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.15555555555555556",
      "ci_low": "0.05405405405405406",
      "ci_high": "0.2653061224489796",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.32432432432432434",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "80.66666666666666",
      "ci_low": "69.49978813559322",
      "ci_high": "90.69767441860465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "60.666666666666664",
      "ci_low": "39.047619047619044",
      "ci_high": "80.45454545454545",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5515632515632515",
      "ci_low": "0.3235449735449735",
      "ci_high": "0.5668097643097644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7637029637029638",
      "ci_low": "0.44798534798534806",
      "ci_high": "0.7848135198135198",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.29333333333333333",
      "ci_low": "0.22807017543859648",
      "ci_high": "0.36087272235251666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5377777777777778",
      "ci_low": "0.4696969696969697",
      "ci_high": "0.6045454545454545",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5454545454545454",
      "ci_low": "0.4473684210526316",
      "ci_high": "0.6446280991735537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.24444444444444444",
      "ci_low": "0.18468468468468469",
      "ci_high": "0.30472103004291845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.3287671232876712",
      "ci_high": "0.47162019930578875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "53.37777777777778",
      "ci_low": "48.17258883248731",
      "ci_high": "58.50015560165975",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "13.377777777777776",
      "ci_low": "2.4685316726671207",
      "ci_high": "24.226903251387785",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.21369465205894514",
      "ci_low": "0.17004444889094011",
      "ci_high": "0.23703862739510048",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3561577534315752",
      "ci_low": "0.28340741481823356",
      "ci_high": "0.3950643789918341",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4507042253521127",
      "ci_low": "0.3333333333333333",
      "ci_high": "0.5689655172413793",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6338028169014085",
      "ci_low": "0.5147058823529411",
      "ci_high": "0.746031746031746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7111111111111111",
      "ci_low": "0.5686274509803921",
      "ci_high": "0.8393035714285711",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.18309859154929578",
      "ci_low": "0.0967741935483871",
      "ci_high": "0.2790783173734609",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.2535211267605634",
      "ci_low": "0.15584415584415584",
      "ci_high": "0.36",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "63.94366197183099",
      "ci_low": "55.15129870129871",
      "ci_high": "72.64778376139186",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "38.59154929577465",
      "ci_low": "22.46570379193965",
      "ci_high": "53.70967741935484",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4298412698412698",
      "ci_low": "0.2561111111111111",
      "ci_high": "0.44037590187590187",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4776014109347443",
      "ci_low": "0.28456790123456793",
      "ci_high": "0.48930655763989095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6578073089700996",
      "ci_low": "0.6028356438579583",
      "ci_high": "0.7105284312066339",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8438538205980066",
      "ci_low": "0.8027199508565969",
      "ci_high": "0.8836496270293989",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7795275590551181",
      "ci_low": "0.7265625",
      "ci_high": "0.8292707123499806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.18604651162790697",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.2317470829391359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.23588039867109634",
      "ci_low": "0.18855218855218855",
      "ci_high": "0.28619557645419713",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "75.28239202657807",
      "ci_low": "71.14580702861953",
      "ci_high": "79.35837467450838",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "51.69435215946844",
      "ci_low": "43.48546130634887",
      "ci_high": "59.72053872053872",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.548301740914744",
      "ci_low": "0.4666071433918518",
      "ci_high": "0.5655581090191384",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6853771761434301",
      "ci_low": "0.5832589292398148",
      "ci_high": "0.7069476362739231",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6538461538461539",
      "ci_low": "0.5365853658536586",
      "ci_high": "0.7634520609318994",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8974358974358975",
      "ci_low": "0.8255813953488372",
      "ci_high": "0.9583333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7285714285714285",
      "ci_low": "0.6124936224489796",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.24358974358974358",
      "ci_low": "0.14752892960462874",
      "ci_high": "0.35",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.2564102564102564",
      "ci_low": "0.15789473684210525",
      "ci_high": "0.36486486486486486",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "78.46153846153845",
      "ci_low": "70.69759366925065",
      "ci_high": "85.80256869772998",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "52.82051282051282",
      "ci_low": "35.570697074010326",
      "ci_high": "69.01234567901234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5159123104203195",
      "ci_low": "0.3739063402907797",
      "ci_high": "0.5601796301524562",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7416239462292094",
      "ci_low": "0.5374903641679958",
      "ci_high": "0.8052582183441558",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.6018518518518519",
      "ci_low": "0.5471698113207547",
      "ci_high": "0.6570605187319885",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.8240740740740741",
      "ci_low": "0.7835341579928539",
      "ci_high": "0.8626229405500219",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7303370786516854",
      "ci_low": "0.67578125",
      "ci_high": "0.7844548105375047",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.2222222222222222",
      "ci_low": "0.17717717717717718",
      "ci_high": "0.26726797552269255",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.2654320987654321",
      "ci_low": "0.21751105379513633",
      "ci_high": "0.31446540880503143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "72.5",
      "ci_low": "68.37416068823565",
      "ci_high": "76.5538719512195",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "45.956790123456784",
      "ci_low": "37.529962707974754",
      "ci_high": "54.08672138170065",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.39320527755125284",
      "ci_low": "0.35413933447537865",
      "ci_high": "0.4131148520925158",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.6208504382388202",
      "ci_low": "0.559167370224282",
      "ci_high": "0.6522866085671303",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3815789473684211",
      "ci_low": "0.2682926829268293",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.7368421052631579",
      "ci_low": "0.6363636363636364",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.5178571428571429",
      "ci_low": "0.38",
      "ci_high": "0.6538461538461539",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.35526315789473684",
      "ci_low": "0.24675324675324675",
      "ci_high": "0.4677572478562676",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.47368421052631576",
      "ci_low": "0.35896739130434785",
      "ci_high": "0.5869651483781918",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "56.44736842105263",
      "ci_low": "47.807866639806605",
      "ci_high": "65.21193829644533",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "9.078947368421053",
      "ci_low": "-9.276036400404449",
      "ci_high": "27.65640219155844",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.22140615948665487",
      "ci_low": "0.12458978311251037",
      "ci_high": "0.23829817755676638",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3653201631529805",
      "ci_low": "0.20557314213564212",
      "ci_high": "0.39319199296866447",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6190476190476191",
      "ci_low": "0.5108695652173914",
      "ci_high": "0.7191011235955056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.780952380952381",
      "ci_low": "0.693069306930693",
      "ci_high": "0.8623881880733945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.7926829268292683",
      "ci_low": "0.691358024691358",
      "ci_high": "0.8809523809523809",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.1619047619047619",
      "ci_low": "0.09259259259259259",
      "ci_high": "0.24000961538461524",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.22857142857142856",
      "ci_low": "0.14166482300884956",
      "ci_high": "0.3218439509059029",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "73.80952380952381",
      "ci_low": "66.3995238095238",
      "ci_high": "80.8108108108108",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "50.95238095238095",
      "ci_low": "35.69974137931034",
      "ci_high": "65.14391534391532",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5107671498194374",
      "ci_low": "0.38266519153619105",
      "ci_high": "0.540111764994806",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5746130435468672",
      "ci_low": "0.43049834047821506",
      "ci_high": "0.6076257356191567",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4105263157894737",
      "ci_low": "0.3511450381679389",
      "ci_high": "0.4696497724852357",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.6070175438596491",
      "ci_low": "0.5480760676156584",
      "ci_high": "0.6654418377063055",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.6763005780346821",
      "ci_low": "0.6012269938650306",
      "ci_high": "0.7471910112359551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.19649122807017544",
      "ci_low": "0.15034965034965034",
      "ci_high": "0.2463768115942029",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3017543859649123",
      "ci_low": "0.24528121159546012",
      "ci_high": "0.3587812857408835",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "60.631578947368425",
      "ci_low": "56.06550922937911",
      "ci_high": "65.13335570469799",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "30.456140350877195",
      "ci_low": "21.814736620720893",
      "ci_high": "38.929797601199404",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.32367967828292854",
      "ci_low": "0.26207503429174434",
      "ci_high": "0.3453264068046811",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.43157290437723805",
      "ci_low": "0.349433379055659",
      "ci_high": "0.46043520907290814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7551020408163265",
      "ci_low": "0.6956521739130435",
      "ci_high": "0.8122275198288923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9306122448979591",
      "ci_low": "0.8966942148760331",
      "ci_high": "0.9610894941634242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8114035087719298",
      "ci_low": "0.7567567567567568",
      "ci_high": "0.863854398319743",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.17551020408163265",
      "ci_low": "0.12652940227474388",
      "ci_high": "0.22608695652173913",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.15017857142857144",
      "ci_high": "0.252",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "86.0",
      "ci_low": "82.43432425055725",
      "ci_high": "89.38867191504191",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "66.0",
      "ci_low": "57.44678937700903",
      "ci_high": "74.21510758050727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.5714178464178464",
      "ci_low": "0.47556384306384314",
      "ci_high": "0.5685184446581505",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7166935700834006",
      "ci_low": "0.5964699048597354",
      "ci_high": "0.713057032283104",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5692307692307692",
      "ci_low": "0.48360655737704916",
      "ci_high": "0.6549356483729966",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7692307692307693",
      "ci_low": "0.6949152542372882",
      "ci_high": "0.8409090909090909",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.74",
      "ci_low": "0.6511627906976745",
      "ci_high": "0.8240766178266178",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.13380281690140844",
      "ci_high": "0.271189725397451",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.3",
      "ci_low": "0.22033898305084745",
      "ci_high": "0.3787878787878788",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "75.76923076923077",
      "ci_low": "70.53815789473684",
      "ci_high": "80.93048101769031",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "45.76923076923077",
      "ci_low": "33.701231987190894",
      "ci_high": "57.89920016406891",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4380217519106408",
      "ci_low": "0.3333241809630698",
      "ci_high": "0.4464432209040053",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5375721500721501",
      "ci_low": "0.40907967663649475",
      "ci_high": "0.5479075892912791",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.15789473684210525",
      "ci_low": "0.078125",
      "ci_high": "0.24358974358974358",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.47368421052631576",
      "ci_low": "0.36",
      "ci_high": "0.5853658536585366",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.18181818181818182",
      "ci_high": "0.48717948717948717",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3157894736842105",
      "ci_low": "0.21621621621621623",
      "ci_high": "0.4225352112676056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.39473684210526316",
      "ci_low": "0.2857142857142857",
      "ci_high": "0.506333223738287",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "32.36842105263158",
      "ci_low": "24.545454545454547",
      "ci_high": "40.394868421052635",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "-7.105263157894736",
      "ci_low": "-21.62170608108108",
      "ci_high": "7.820783132530116",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.1596045197740113",
      "ci_low": "0.057908931934355666",
      "ci_high": "0.1556383447061413",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.20471014492753623",
      "ci_low": "0.07427449965493443",
      "ci_high": "0.1996230942970073",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.20408163265306123",
      "ci_low": "0.12631578947368421",
      "ci_high": "0.2894798327594687",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4489795918367347",
      "ci_low": "0.3490566037735849",
      "ci_high": "0.550561797752809",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.45454545454545453",
      "ci_low": "0.3055555555555556",
      "ci_high": "0.6052631578947368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.24489795918367346",
      "ci_low": "0.1619047619047619",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.3877551020408163",
      "ci_low": "0.28888888888888886",
      "ci_high": "0.49",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "36.224489795918366",
      "ci_low": "28.8886292834891",
      "ci_high": "44.14160219213411",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "-2.5510204081632653",
      "ci_low": "-17.448979591836736",
      "ci_high": "12.088655462184862",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.14779183081069872",
      "ci_low": "0.08113157282968603",
      "ci_high": "0.16378640135299402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.17028189202102245",
      "ci_low": "0.09347768173855131",
      "ci_high": "0.18871041895018875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5794392523364486",
      "ci_low": "0.48514851485148514",
      "ci_high": "0.6698113207547169",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7850467289719626",
      "ci_low": "0.7064220183486238",
      "ci_high": "0.8571428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7380952380952381",
      "ci_low": "0.6395187338501293",
      "ci_high": "0.8300127118644065",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.205607476635514",
      "ci_low": "0.1308411214953271",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2897196261682243",
      "ci_low": "0.20407667921537548",
      "ci_high": "0.3786407766990291",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "71.12149532710279",
      "ci_low": "64.22217125382264",
      "ci_high": "77.80952380952381",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "42.149532710280376",
      "ci_low": "27.63631476050831",
      "ci_high": "56.272771836007124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5098466505246166",
      "ci_low": "0.3329161163695062",
      "ci_high": "0.4811956122973072",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5898225957049487",
      "ci_low": "0.3851382522706052",
      "ci_high": "0.5566772769713946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.39097744360902253",
      "ci_low": "0.30767716535433076",
      "ci_high": "0.47580784204671855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6390977443609023",
      "ci_low": "0.5555555555555556",
      "ci_high": "0.7196969696969697",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.611764705882353",
      "ci_low": "0.5046428571428573",
      "ci_high": "0.7142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.24812030075187969",
      "ci_low": "0.17518248175182483",
      "ci_high": "0.32608695652173914",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3383458646616541",
      "ci_low": "0.2569444444444444",
      "ci_high": "0.4227655917498772",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "61.42857142857143",
      "ci_low": "55.24806966369252",
      "ci_high": "67.42424242424242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "27.593984962406015",
      "ci_low": "15.211267605633802",
      "ci_high": "39.687560562015506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.35121693121693115",
      "ci_low": "0.2299664953102453",
      "ci_high": "0.34595628908128906",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.4300615484288954",
      "ci_low": "0.2815916269105045",
      "ci_high": "0.4236199458138234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Flash-Lite HIGH",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6503496503496503",
      "ci_low": "0.6017073642589098",
      "ci_high": "0.6977912871355494",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7482517482517482",
      "ci_low": "0.7034266356168403",
      "ci_high": "0.7906976744186046",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8691588785046729",
      "ci_low": "0.8292008880447617",
      "ci_high": "0.9060791360474888",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.0979020979020979",
      "ci_low": "0.06994773361353672",
      "ci_high": "0.12809276865456637",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.1282051282051282",
      "ci_low": "0.09579187720788802",
      "ci_high": "0.16312056737588654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "75.96736596736596",
      "ci_low": "72.47705350252936",
      "ci_high": "79.38920906200319",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "63.14685314685315",
      "ci_low": "57.036066089143496",
      "ci_high": "68.88356553961209",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5693708656290797",
      "ci_low": "0.4885538818334084",
      "ci_high": "0.5733599868573677",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7071218815070828",
      "ci_low": "0.6067524016318138",
      "ci_high": "0.7120761127099566",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.44598337950138506",
      "ci_low": "0.39322749732905987",
      "ci_high": "0.498567335243553",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5761772853185596",
      "ci_low": "0.5220994475138122",
      "ci_high": "0.6293103448275862",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.7740384615384616",
      "ci_low": "0.7162790697674418",
      "ci_high": "0.8287068318318318",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.13019390581717452",
      "ci_low": "0.09710886208476019",
      "ci_high": "0.1657612328528249",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.19390581717451524",
      "ci_low": "0.15363881401617252",
      "ci_high": "0.23697916666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "63.739612188365655",
      "ci_low": "59.77650492749984",
      "ci_high": "67.67960816691873",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "44.34903047091413",
      "ci_low": "37.45969709178815",
      "ci_high": "51.06304965001876",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3922731053912909",
      "ci_low": "0.31622015203762194",
      "ci_high": "0.4006330134517594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.458760072406764",
      "ci_low": "0.3698167879762019",
      "ci_high": "0.4685369140368034",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7386666666666667",
      "ci_low": "0.6916395027521984",
      "ci_high": "0.7849506798187149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8453333333333334",
      "ci_low": "0.8061776354695136",
      "ci_high": "0.8841322275825756",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8738170347003155",
      "ci_low": "0.8360128617363344",
      "ci_high": "0.909375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.10666666666666667",
      "ci_low": "0.07646722047597666",
      "ci_high": "0.13920544447640965",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.13066666666666665",
      "ci_low": "0.09714210723659543",
      "ci_high": "0.1671018276762402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "85.28",
      "ci_low": "82.4542960380829",
      "ci_high": "88.03664921465968",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "72.21333333333332",
      "ci_low": "66.40649219911384",
      "ci_high": "77.67384049151248",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.6917675788264023",
      "ci_low": "0.6092684192932306",
      "ci_high": "0.6934404355584225",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7538492846185153",
      "ci_low": "0.6639463543580076",
      "ci_high": "0.7556722695187938",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.22988505747126436",
      "ci_low": "0.16867294760437096",
      "ci_high": "0.2947383632491633",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3735632183908046",
      "ci_low": "0.3016759776536313",
      "ci_high": "0.449438202247191",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.6153846153846154",
      "ci_low": "0.5",
      "ci_high": "0.734375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.14367816091954022",
      "ci_low": "0.09259259259259259",
      "ci_high": "0.19753086419753085",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.19540229885057472",
      "ci_low": "0.13736263736263737",
      "ci_high": "0.2561599445812808",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "39.770114942528735",
      "ci_low": "34.17142093200917",
      "ci_high": "45.4914088801372",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "20.229885057471265",
      "ci_low": "11.140846605440595",
      "ci_high": "29.26986579275905",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.23909191583610187",
      "ci_low": "0.13719663411523877",
      "ci_high": "0.22759589247961343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2670377241805813",
      "ci_low": "0.15323260433650043",
      "ci_high": "0.2541980097824254",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5083333333333333",
      "ci_low": "0.4435146443514644",
      "ci_high": "0.572",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6083333333333333",
      "ci_low": "0.5450368050977807",
      "ci_high": "0.6709429824561403",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.8356164383561644",
      "ci_low": "0.7741880247459125",
      "ci_high": "0.891724048889654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.1",
      "ci_low": "0.06493506493506493",
      "ci_high": "0.13924310022826308",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.175",
      "ci_low": "0.12704918032786885",
      "ci_high": "0.22666801619433194",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "69.16666666666667",
      "ci_low": "64.64127390250472",
      "ci_high": "73.55857764544206",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "51.66666666666667",
      "ci_low": "43.267348923252065",
      "ci_high": "59.43331219615303",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5349987267634326",
      "ci_low": "0.403278626602156",
      "ci_high": "0.5148592693144843",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5756315414543262",
      "ci_low": "0.4339073830529527",
      "ci_high": "0.5539625049586224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.625",
      "ci_low": "0.5748651987247069",
      "ci_high": "0.6742424242424242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7425",
      "ci_low": "0.6974342105263157",
      "ci_high": "0.7862237156180149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.8417508417508418",
      "ci_low": "0.7979443105341033",
      "ci_high": "0.8819708783153405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1175",
      "ci_low": "0.08730109250682816",
      "ci_high": "0.15",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.15",
      "ci_low": "0.11491442542787286",
      "ci_high": "0.1876433031047829",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "73.675",
      "ci_low": "69.91807727995759",
      "ci_high": "77.2959528405957",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "58.675",
      "ci_low": "52.14277408637874",
      "ci_high": "64.90440197666679",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5250346685202693",
      "ci_low": "0.4828035179866669",
      "ci_high": "0.5469169879771919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6956709357893567",
      "ci_low": "0.6397146613323337",
      "ci_high": "0.7246650090697794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.48717948717948717",
      "ci_low": "0.43478260869565216",
      "ci_high": "0.5398936170212766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5948717948717949",
      "ci_low": "0.5424164524421594",
      "ci_high": "0.6466165413533834",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.8189655172413793",
      "ci_low": "0.7672413793103449",
      "ci_high": "0.8674711667181547",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.1076923076923077",
      "ci_low": "0.07780899775856549",
      "ci_high": "0.13994910941475827",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.16666666666666666",
      "ci_low": "0.129576370137926",
      "ci_high": "0.2074468085106383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "67.0",
      "ci_low": "63.24045222010888",
      "ci_high": "70.72845673702551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "50.33333333333333",
      "ci_low": "43.93546676096181",
      "ci_high": "56.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4834808090277646",
      "ci_low": "0.4369546180370484",
      "ci_high": "0.5089072946184343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.557862471955113",
      "ci_low": "0.5041784054273635",
      "ci_high": "0.5872007245597318",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6943521594684385",
      "ci_low": "0.6416924214294448",
      "ci_high": "0.7448996712779071",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7906976744186046",
      "ci_low": "0.7439446366782007",
      "ci_high": "0.83544405879951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8781512605042017",
      "ci_low": "0.8353897069760523",
      "ci_high": "0.9173553719008265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.09634551495016612",
      "ci_low": "0.06498077734539828",
      "ci_high": "0.13071895424836602",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.13953488372093023",
      "ci_low": "0.10135024730612967",
      "ci_high": "0.17981072555205047",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "77.30897009966777",
      "ci_low": "73.14670745920746",
      "ci_high": "81.29040717347168",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "63.355481727574755",
      "ci_low": "56.37575667851033",
      "ci_high": "70.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6747916622061628",
      "ci_low": "0.5825311733675967",
      "ci_high": "0.6928870944656818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7254010368716252",
      "ci_low": "0.6262210113701665",
      "ci_high": "0.7448536265506079",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.37407407407407406",
      "ci_low": "0.3109517475946586",
      "ci_high": "0.44068252232549654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.46296296296296297",
      "ci_low": "0.39642694805194806",
      "ci_high": "0.5286897022800074",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.808",
      "ci_low": "0.7323915936564268",
      "ci_high": "0.8761061946902655",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.08888888888888889",
      "ci_low": "0.05574912891986063",
      "ci_high": "0.125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.15925925925925927",
      "ci_low": "0.11235853281224556",
      "ci_high": "0.21070234113712374",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "60.77777777777777",
      "ci_low": "56.12161625333409",
      "ci_high": "65.49824435960623",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "44.85185185185185",
      "ci_low": "36.59852854185152",
      "ci_high": "52.907833655706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3654038599859033",
      "ci_low": "0.31945661070449216",
      "ci_high": "0.3972301497199399",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.44660471776054855",
      "ci_low": "0.3904469686388237",
      "ci_high": "0.48550351632437105",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.593607305936073",
      "ci_low": "0.5235849056603774",
      "ci_high": "0.6594850513573",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7579908675799086",
      "ci_low": "0.7013574660633484",
      "ci_high": "0.8135612480739599",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7831325301204819",
      "ci_low": "0.7151898734177216",
      "ci_high": "0.8445977421520899",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1643835616438356",
      "ci_low": "0.11666479820627804",
      "ci_high": "0.21524824597163514",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.182648401826484",
      "ci_low": "0.13333333333333333",
      "ci_high": "0.23660714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "72.69406392694064",
      "ci_low": "67.78242677824268",
      "ci_high": "77.42097876494996",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "54.42922374429224",
      "ci_low": "45.42185979971388",
      "ci_high": "62.73594598087361",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4713145565209549",
      "ci_low": "0.398615697767336",
      "ci_high": "0.4990928341876644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6427016679831203",
      "ci_low": "0.543566860591822",
      "ci_high": "0.6805811375286333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6917808219178082",
      "ci_low": "0.6382211171106922",
      "ci_high": "0.7428571428571429",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7876712328767124",
      "ci_low": "0.7397260273972602",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8782608695652174",
      "ci_low": "0.8346960557753319",
      "ci_high": "0.9181034482758621",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.0958904109589041",
      "ci_low": "0.06418918918918919",
      "ci_high": "0.13114754098360656",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.1404109589041096",
      "ci_low": "0.10150375939849623",
      "ci_high": "0.1812363956742562",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "77.05479452054794",
      "ci_low": "72.85691169671753",
      "ci_high": "81.11906608749612",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "63.013698630136986",
      "ci_low": "55.93856655290103",
      "ci_high": "69.81044079360804",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6780522123137518",
      "ci_low": "0.5748255255959133",
      "ci_high": "0.6929175371645308",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7276657888245142",
      "ci_low": "0.6168859299078093",
      "ci_high": "0.7436188203716916",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.6157894736842106",
      "ci_low": "0.5384523809523809",
      "ci_high": "0.6898154337492572",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7526315789473684",
      "ci_low": "0.6827956989247311",
      "ci_high": "0.8192812918340026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8181818181818182",
      "ci_low": "0.75",
      "ci_high": "0.8785714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.1368421052631579",
      "ci_low": "0.09045226130653267",
      "ci_high": "0.18811881188118812",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.1631578947368421",
      "ci_low": "0.11290322580645161",
      "ci_high": "0.21818181818181817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "73.73684210526315",
      "ci_low": "67.92482951445719",
      "ci_high": "79.30514705882354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "57.42105263157895",
      "ci_low": "48.090225838494796",
      "ci_high": "66.19883040935673",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4463897807030101",
      "ci_low": "0.3625681531784548",
      "ci_high": "0.4618479658249916",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5661528925989396",
      "ci_low": "0.45984253573852796",
      "ci_high": "0.5857583956804772",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3146067415730337",
      "ci_low": "0.21839080459770116",
      "ci_high": "0.4144144144144144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.48314606741573035",
      "ci_low": "0.37362637362637363",
      "ci_high": "0.5876288659793815",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6511627906976745",
      "ci_low": "0.5",
      "ci_high": "0.8048780487804879",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.16853932584269662",
      "ci_low": "0.08695652173913043",
      "ci_high": "0.25842696629213485",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2247191011235955",
      "ci_low": "0.13043478260869565",
      "ci_high": "0.3287671232876712",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "55.84269662921348",
      "ci_low": "48.57142857142857",
      "ci_high": "63.16471133065761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "33.37078651685393",
      "ci_low": "18.75",
      "ci_high": "47.40029411764706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.26886446886446885",
      "ci_low": "0.14491275037598567",
      "ci_high": "0.28440038543328017",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3666333666333666",
      "ci_low": "0.19760829596725318",
      "ci_high": "0.3878187074090184",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "9933",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "41.42857142857143",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "51.66666666666667",
      "ci_low": "16.666666666666668",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.4129353233830846",
      "ci_low": "0.3427230046948357",
      "ci_high": "0.4869109947643979",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5074626865671642",
      "ci_low": "0.4362719023779725",
      "ci_high": "0.5789533222195761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.8137254901960784",
      "ci_low": "0.7346938775510204",
      "ci_high": "0.8861788617886179",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.0945273631840796",
      "ci_low": "0.056870691512279185",
      "ci_high": "0.13596893083494072",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.15422885572139303",
      "ci_low": "0.10328574709124312",
      "ci_high": "0.2079207920792079",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "64.12935323383084",
      "ci_low": "59.03834541062802",
      "ci_high": "69.07219009303495",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "48.70646766169155",
      "ci_low": "39.666545893719814",
      "ci_high": "57.34905410290397",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.40944205146991525",
      "ci_low": "0.3252832781813573",
      "ci_high": "0.4356984952578367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5004291740187854",
      "ci_low": "0.3975684511105477",
      "ci_high": "0.5325203830929115",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "10.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "-90.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.28125",
      "ci_high": "0.5172515212981743",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.5857142857142857",
      "ci_low": "0.4666666666666667",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6829268292682927",
      "ci_low": "0.5263157894736842",
      "ci_high": "0.8333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.18571428571428572",
      "ci_low": "0.09333333333333334",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.10665957446808512",
      "ci_high": "0.3055555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "58.57142857142857",
      "ci_low": "49.50819672131147",
      "ci_high": "67.40034482758621",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "38.57142857142857",
      "ci_low": "22.878246753246756",
      "ci_high": "53.125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2827708006279435",
      "ci_low": "0.19731724075349683",
      "ci_high": "0.33223931239162335",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.565541601255887",
      "ci_low": "0.39463448150699365",
      "ci_high": "0.6644786247832466",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.44680851063829785",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8444444444444444",
      "ci_low": "0.7307692307692307",
      "ci_high": "0.9428571428571428",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7105263157894737",
      "ci_low": "0.5526315789473685",
      "ci_high": "0.851063829787234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.24444444444444444",
      "ci_low": "0.125",
      "ci_high": "0.375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.26666666666666666",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.40425531914893614",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "74.22222222222223",
      "ci_low": "63.33333333333333",
      "ci_high": "84.44444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "47.55555555555556",
      "ci_low": "25.713474025974026",
      "ci_high": "68.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.41954887218045117",
      "ci_low": "0.23114628676955545",
      "ci_high": "0.4365726342371079",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6131868131868132",
      "ci_low": "0.33782918835550413",
      "ci_high": "0.6380676961926962",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.37333333333333335",
      "ci_low": "0.30195893451720307",
      "ci_high": "0.44723671098106565",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.4666666666666667",
      "ci_low": "0.39285714285714285",
      "ci_high": "0.5401812161475964",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.8",
      "ci_low": "0.7156862745098039",
      "ci_high": "0.8761904761904762",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.09333333333333334",
      "ci_low": "0.05583756345177665",
      "ci_high": "0.1341991341991342",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.16444444444444445",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.22169864680269785",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "60.75555555555556",
      "ci_low": "55.658221496090555",
      "ci_high": "65.98216193528694",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "44.31111111111111",
      "ci_low": "35.13746378561082",
      "ci_high": "53.23285251824295",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.38284860698512013",
      "ci_low": "0.323555213041656",
      "ci_high": "0.41834784139471637",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.45374649716754983",
      "ci_low": "0.3834728450864072",
      "ci_high": "0.4958196638752194",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4788732394366197",
      "ci_low": "0.36231884057971014",
      "ci_high": "0.5942028985507246",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5633802816901409",
      "ci_low": "0.4461538461538462",
      "ci_high": "0.6779661016949152",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.85",
      "ci_low": "0.7272727272727273",
      "ci_high": "0.9512195121951219",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.08450704225352113",
      "ci_low": "0.02702702702702703",
      "ci_high": "0.15625941265060228",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.1267605633802817",
      "ci_low": "0.05555555555555555",
      "ci_high": "0.21052631578947367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "66.7605633802817",
      "ci_low": "58.0766129032258",
      "ci_high": "75.07936507936508",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "54.08450704225352",
      "ci_low": "39.835997267759566",
      "ci_high": "67.0698729582577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4965136054421769",
      "ci_low": "0.31144605394605396",
      "ci_high": "0.5134834312512885",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5149029982363316",
      "ci_low": "0.32298109298109307",
      "ci_high": "0.5325013361124472",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6943521594684385",
      "ci_low": "0.6416924214294448",
      "ci_high": "0.7448996712779071",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7906976744186046",
      "ci_low": "0.7439446366782007",
      "ci_high": "0.83544405879951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8781512605042017",
      "ci_low": "0.8353897069760523",
      "ci_high": "0.9173553719008265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.09634551495016612",
      "ci_low": "0.06498077734539828",
      "ci_high": "0.13071895424836602",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.13953488372093023",
      "ci_low": "0.10135024730612967",
      "ci_high": "0.17981072555205047",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "77.30897009966777",
      "ci_low": "73.14670745920746",
      "ci_high": "81.29040717347168",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "63.355481727574755",
      "ci_low": "56.37575667851033",
      "ci_high": "70.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6747916622061628",
      "ci_low": "0.5825311733675967",
      "ci_high": "0.6928870944656818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7254010368716252",
      "ci_low": "0.6262210113701665",
      "ci_high": "0.7448536265506079",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7435897435897436",
      "ci_low": "0.6351351351351351",
      "ci_high": "0.8421052631578947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8589743589743589",
      "ci_low": "0.7763157894736842",
      "ci_high": "0.9324362283631946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8656716417910447",
      "ci_low": "0.7735849056603774",
      "ci_high": "0.9411764705882353",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.11538461538461539",
      "ci_low": "0.05084745762711865",
      "ci_high": "0.1917808219178082",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.14102564102564102",
      "ci_low": "0.07041635779589821",
      "ci_high": "0.2261904761904762",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "83.07692307692308",
      "ci_low": "75.57103174603175",
      "ci_high": "89.85714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "68.97435897435898",
      "ci_low": "54.48694029850746",
      "ci_high": "81.60032258064516",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5598262032085561",
      "ci_low": "0.41684905780494025",
      "ci_high": "0.5974042863701954",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7697610294117647",
      "ci_low": "0.5731674544817927",
      "ci_high": "0.8214308937590187",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.6728395061728395",
      "ci_low": "0.6191948214414286",
      "ci_high": "0.7256944444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7654320987654321",
      "ci_low": "0.717391304347826",
      "ci_high": "0.8117293028322441",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.8790322580645161",
      "ci_low": "0.8352919145186165",
      "ci_high": "0.9181826630618453",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.09259259259259259",
      "ci_low": "0.0625",
      "ci_high": "0.1261003672293995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.11728395061728394",
      "ci_low": "0.08231707317073171",
      "ci_high": "0.15540746539222144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "76.9753086419753",
      "ci_low": "72.92993630573248",
      "ci_high": "80.90361445783132",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "65.24691358024691",
      "ci_low": "58.41202759963353",
      "ci_high": "71.76123712743274",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5778212733813252",
      "ci_low": "0.5153238895506652",
      "ci_high": "0.6001641279245447",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7298795032185161",
      "ci_low": "0.6509354394324193",
      "ci_high": "0.7581020563257408",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.42105263157894735",
      "ci_low": "0.3",
      "ci_high": "0.5405516171909613",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6447368421052632",
      "ci_low": "0.5277701465201465",
      "ci_high": "0.7571428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6530612244897959",
      "ci_low": "0.5094253859348199",
      "ci_high": "0.7857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2236842105263158",
      "ci_low": "0.13333333333333333",
      "ci_high": "0.32222222222222224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2894736842105263",
      "ci_low": "0.18518518518518517",
      "ci_high": "0.40229885057471265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "59.60526315789473",
      "ci_low": "50.270270270270274",
      "ci_high": "68.53341463414634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "30.657894736842106",
      "ci_low": "12.434701492537314",
      "ci_high": "47.80916754478397",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2977832512315271",
      "ci_low": "0.17351796909986567",
      "ci_high": "0.3246398716226302",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.43178571428571433",
      "ci_low": "0.2516010551948052",
      "ci_high": "0.47072781385281376",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.580952380952381",
      "ci_low": "0.475",
      "ci_high": "0.6853932584269663",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6952380952380952",
      "ci_low": "0.5929203539823009",
      "ci_high": "0.7943987817089452",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8356164383561644",
      "ci_low": "0.7384615384615385",
      "ci_high": "0.9210526315789473",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.11428571428571428",
      "ci_low": "0.053763440860215055",
      "ci_high": "0.18447393414943003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.1619047619047619",
      "ci_low": "0.0888818565400844",
      "ci_high": "0.24490296118447372",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "72.85714285714286",
      "ci_low": "65.8331043956044",
      "ci_high": "79.87345727848101",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "56.66666666666667",
      "ci_low": "43.18122009569379",
      "ci_high": "69.10127407704654",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5553839838322597",
      "ci_low": "0.39146388179096087",
      "ci_high": "0.5703936970455734",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6710889804639805",
      "ci_low": "0.4730188571640777",
      "ci_high": "0.689225717263401",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.45263157894736844",
      "ci_low": "0.3919376332622601",
      "ci_high": "0.5124577481855636",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5578947368421052",
      "ci_low": "0.4966216216216216",
      "ci_high": "0.6181818181818182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.8113207547169812",
      "ci_low": "0.7485705931495406",
      "ci_high": "0.8692819741532977",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.10526315789473684",
      "ci_low": "0.0711743772241993",
      "ci_high": "0.14145524551320818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.16842105263157894",
      "ci_low": "0.12624584717607973",
      "ci_high": "0.21400814105417318",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "64.84210526315789",
      "ci_low": "60.37536931384036",
      "ci_high": "69.19016673570836",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "48.0",
      "ci_low": "40.78665755919854",
      "ci_high": "54.91805931349143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4403752862123052",
      "ci_low": "0.37129940246965343",
      "ci_high": "0.4632153407410484",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.47425030822863634",
      "ci_low": "0.399860894967319",
      "ci_high": "0.4988472900288214",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8",
      "ci_low": "0.746720363028998",
      "ci_high": "0.8512412055669913",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8938775510204081",
      "ci_low": "0.8493123505109806",
      "ci_high": "0.9356223175965666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8949771689497716",
      "ci_low": "0.8530612244897959",
      "ci_high": "0.9339622641509434",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.09387755102040816",
      "ci_low": "0.05859375",
      "ci_high": "0.13158165475416933",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.11020408163265306",
      "ci_low": "0.07199924242424242",
      "ci_high": "0.15126050420168066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "88.77551020408163",
      "ci_low": "85.66659292035398",
      "ci_high": "91.73236480770483",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "77.75510204081633",
      "ci_low": "71.11505102040816",
      "ci_high": "83.99159663865547",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7142061642061642",
      "ci_low": "0.5822732846372024",
      "ci_high": "0.695857211341837",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7989424887729972",
      "ci_low": "0.6513565556958536",
      "ci_high": "0.778416541501038",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6230769230769231",
      "ci_low": "0.5371862565044383",
      "ci_high": "0.7086614173228346",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7538461538461538",
      "ci_low": "0.6759243295019157",
      "ci_high": "0.8297872340425532",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.826530612244898",
      "ci_low": "0.75",
      "ci_high": "0.8977272727272727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.13076923076923078",
      "ci_low": "0.07627118644067797",
      "ci_high": "0.1896551724137931",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.16923076923076924",
      "ci_low": "0.10526315789473684",
      "ci_high": "0.23741744309470444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "78.6923076923077",
      "ci_low": "73.21428571428571",
      "ci_high": "84.07094962559565",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "61.769230769230774",
      "ci_low": "50.93178562902841",
      "ci_high": "71.98122209419559",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5803902116402117",
      "ci_low": "0.43799546807359313",
      "ci_high": "0.5790610166514771",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6331529581529581",
      "ci_low": "0.4778132378984651",
      "ci_high": "0.6317029272561568",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14473684210526316",
      "ci_high": "0.33783783783783783",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3684210526315789",
      "ci_low": "0.2602739726027397",
      "ci_high": "0.4794520547945205",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.6428571428571429",
      "ci_low": "0.4583333333333333",
      "ci_high": "0.8181818181818182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.13157894736842105",
      "ci_low": "0.060596933187294644",
      "ci_high": "0.2112676056338028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.17105263157894737",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.25925925925925924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "38.81578947368421",
      "ci_low": "30.499845679012346",
      "ci_high": "47.70533305665075",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "21.710526315789473",
      "ci_low": "8.181818181818182",
      "ci_high": "35.422508214676874",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.2759259259259259",
      "ci_low": "0.11724046015712682",
      "ci_high": "0.25044091710758376",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3239130434782609",
      "ci_low": "0.1376301054018445",
      "ci_high": "0.2939958592132505",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.22448979591836735",
      "ci_low": "0.14457496653279786",
      "ci_high": "0.31314616755793206",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.37755102040816324",
      "ci_low": "0.2803738317757009",
      "ci_high": "0.47959183673469385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.5945945945945946",
      "ci_low": "0.4358974358974359",
      "ci_high": "0.75",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.15306122448979592",
      "ci_low": "0.08602150537634409",
      "ci_high": "0.22772277227722773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.1346153846153846",
      "ci_high": "0.2990887850467286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "40.51020408163265",
      "ci_low": "33.106559841910816",
      "ci_high": "48.421637426900574",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "19.081632653061224",
      "ci_low": "6.736788399570354",
      "ci_high": "31.386138613861387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.18898225957049486",
      "ci_low": "0.10469771241830067",
      "ci_high": "0.20046724245253655",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.20952380952380947",
      "ci_low": "0.11607789855072465",
      "ci_high": "0.22225716011042096",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5981308411214953",
      "ci_low": "0.5042735042735043",
      "ci_high": "0.6885245901639344",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6822429906542056",
      "ci_low": "0.5904761904761905",
      "ci_high": "0.7663551401869159",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.8767123287671232",
      "ci_low": "0.7966101694915254",
      "ci_high": "0.9473684210526315",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.08411214953271028",
      "ci_low": "0.03571428571428571",
      "ci_high": "0.14130434782608695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.14018691588785046",
      "ci_low": "0.07563025210084033",
      "ci_high": "0.2087987012987012",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "72.80373831775701",
      "ci_low": "65.77971765264158",
      "ci_high": "79.22332189487781",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "58.78504672897196",
      "ci_low": "47.22639866757514",
      "ci_high": "69.70880787901419",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5605442176870749",
      "ci_low": "0.36844954648526074",
      "ci_high": "0.5252251317193544",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6154995331465919",
      "ci_low": "0.40457205104263927",
      "ci_high": "0.5767177916918402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.43609022556390975",
      "ci_low": "0.3515625",
      "ci_high": "0.5220604502863092",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5488721804511278",
      "ci_low": "0.46218487394957986",
      "ci_high": "0.6349206349206349",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.7945205479452054",
      "ci_low": "0.6973684210526315",
      "ci_high": "0.8815982972136219",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.11278195488721804",
      "ci_low": "0.06299074940807224",
      "ci_high": "0.1680016806722689",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.20300751879699247",
      "ci_low": "0.1366906474820144",
      "ci_high": "0.2765957446808511",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "66.24060150375941",
      "ci_low": "60.294063975955346",
      "ci_high": "71.95488721804512",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "45.93984962406015",
      "ci_low": "35.03811302103356",
      "ci_high": "56.209738514174",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.46521578421578424",
      "ci_low": "0.30725842490842487",
      "ci_high": "0.4542333638583639",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.47470998389365743",
      "ci_low": "0.3135290050085969",
      "ci_high": "0.46350343250853454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.7 MEDIUM media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6783216783216783",
      "ci_low": "0.631336405529954",
      "ci_high": "0.7246393861892583",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7575757575757576",
      "ci_low": "0.7136038186157518",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8953846153846153",
      "ci_low": "0.8610271903323263",
      "ci_high": "0.926984126984127",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.07925407925407925",
      "ci_low": "0.055023325358851675",
      "ci_high": "0.10552989667438312",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.1048951048951049",
      "ci_low": "0.07586206896551724",
      "ci_high": "0.1358590827513136",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "77.92540792540792",
      "ci_low": "74.59520633079218",
      "ci_high": "81.1857732781651",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "67.43589743589743",
      "ci_low": "61.884954844006565",
      "ci_high": "72.82462650917448",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6164839736143914",
      "ci_low": "0.5357398042384097",
      "ci_high": "0.6193322491825886",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7159168725844545",
      "ci_low": "0.6221494500833143",
      "ci_high": "0.7192245474378447",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.43490304709141275",
      "ci_low": "0.3808187961364721",
      "ci_high": "0.488890423572744",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5457063711911357",
      "ci_low": "0.4913282195187525",
      "ci_high": "0.5988712048137431",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.7969543147208121",
      "ci_low": "0.7357512953367875",
      "ci_high": "0.8523829039812646",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.11080332409972299",
      "ci_low": "0.07926605750661612",
      "ci_high": "0.14641059061551626",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.1772853185595568",
      "ci_low": "0.13829630788485608",
      "ci_high": "0.21957671957671956",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "63.57340720221607",
      "ci_low": "59.59336441893831",
      "ci_high": "67.52221512635707",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "45.84487534626039",
      "ci_low": "39.04487697557472",
      "ci_high": "52.43041260289134",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.41804116678889525",
      "ci_low": "0.3363835465957599",
      "ci_high": "0.4298271630594536",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4534683843133779",
      "ci_low": "0.3648906268157397",
      "ci_high": "0.4662531938272039",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7626666666666667",
      "ci_low": "0.7164128680919726",
      "ci_high": "0.8079118644067796",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8426666666666667",
      "ci_low": "0.802138331945875",
      "ci_high": "0.8820375335120644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.9050632911392406",
      "ci_low": "0.8706624605678234",
      "ci_high": "0.9367088607594937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.08",
      "ci_low": "0.05316225868043278",
      "ci_high": "0.109375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.09333333333333334",
      "ci_low": "0.06417112299465241",
      "ci_high": "0.12430939226519337",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "86.93333333333334",
      "ci_low": "84.20654412058693",
      "ci_high": "89.5544206489889",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "77.6",
      "ci_low": "72.378339208774",
      "ci_high": "82.48603351955308",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7303325857542724",
      "ci_low": "0.6490451096466334",
      "ci_high": "0.7319890034241242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7771487771487771",
      "ci_low": "0.6906505653932123",
      "ci_high": "0.7789113754384911",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2413793103448276",
      "ci_low": "0.17901234567901234",
      "ci_high": "0.3072639324158605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3390804597701149",
      "ci_low": "0.269836811128946",
      "ci_high": "0.4114285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.711864406779661",
      "ci_low": "0.5909090909090909",
      "ci_high": "0.8245614035087719",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.09770114942528736",
      "ci_low": "0.055865921787709494",
      "ci_high": "0.14450867052023122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1781609195402299",
      "ci_low": "0.12209302325581395",
      "ci_high": "0.23837209302325582",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "41.206896551724135",
      "ci_low": "35.53065685832264",
      "ci_high": "46.98717948717949",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "23.39080459770115",
      "ci_low": "14.057065637065637",
      "ci_high": "32.228739495798315",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.294006919006919",
      "ci_low": "0.1657100075850076",
      "ci_high": "0.2729249835132188",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2978251906823335",
      "ci_low": "0.16786208560559213",
      "ci_high": "0.2764694638185852",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.49583333333333335",
      "ci_low": "0.4297872340425532",
      "ci_high": "0.5603448275862069",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6083333333333333",
      "ci_low": "0.5447154471544715",
      "ci_high": "0.6708860759493671",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.815068493150685",
      "ci_low": "0.7481481481481481",
      "ci_high": "0.875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.1125",
      "ci_low": "0.07441764366711513",
      "ci_high": "0.15451282444308187",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.17916666666666667",
      "ci_low": "0.13022877670637173",
      "ci_high": "0.23333333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "68.79166666666666",
      "ci_low": "64.24114621117194",
      "ci_high": "73.22320714476139",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "50.875",
      "ci_low": "42.378894973013686",
      "ci_high": "58.77637130801688",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5312174100409395",
      "ci_low": "0.39787418300653593",
      "ci_high": "0.5096462758997816",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5715630361199983",
      "ci_low": "0.4280924753867792",
      "ci_high": "0.5483535879934358",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.66",
      "ci_low": "0.6116745108668868",
      "ci_high": "0.7079213704080688",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.76",
      "ci_low": "0.7164547515586625",
      "ci_high": "0.8026370483810477",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.868421052631579",
      "ci_low": "0.8278145695364238",
      "ci_high": "0.9057239057239057",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1",
      "ci_low": "0.07105263157894737",
      "ci_high": "0.13101604278074866",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1275",
      "ci_low": "0.09571788413098237",
      "ci_high": "0.1618798955613577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "75.975",
      "ci_low": "72.36087046204621",
      "ci_high": "79.50257243195786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "63.224999999999994",
      "ci_low": "56.97205914687775",
      "ci_high": "69.15001213592234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.6194118479870407",
      "ci_low": "0.5731193928964311",
      "ci_high": "0.6434837705102915",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7278089213847727",
      "ci_low": "0.6734152866533064",
      "ci_high": "0.7560934303495924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4717948717948718",
      "ci_low": "0.4190205308135899",
      "ci_high": "0.524632910193255",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.558974358974359",
      "ci_low": "0.5064926675391791",
      "ci_high": "0.6112663220629929",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.8440366972477065",
      "ci_low": "0.7923655395200537",
      "ci_high": "0.8909952606635071",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.08717948717948718",
      "ci_low": "0.05970149253731343",
      "ci_high": "0.11716674472070844",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.14871794871794872",
      "ci_low": "0.11232876712328767",
      "ci_high": "0.18812406439582943",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "66.64102564102565",
      "ci_low": "62.85339371124221",
      "ci_high": "70.30626159554731",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "51.769230769230774",
      "ci_low": "45.50840555288075",
      "ci_high": "57.80945582103297",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5056045771004524",
      "ci_low": "0.45544501401993476",
      "ci_high": "0.5327076861406577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5444972368774103",
      "ci_low": "0.49047924586762204",
      "ci_high": "0.5736852004591699",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.707641196013289",
      "ci_low": "0.6559476886906614",
      "ci_high": "0.7580085524050052",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7873754152823921",
      "ci_low": "0.740484429065744",
      "ci_high": "0.8316831683168316",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8987341772151899",
      "ci_low": "0.8577981651376146",
      "ci_high": "0.9358974358974359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.07973421926910298",
      "ci_low": "0.050292941825199895",
      "ci_high": "0.11224489795918367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.132890365448505",
      "ci_low": "0.09570591018146184",
      "ci_high": "0.17307692307692307",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "78.63787375415282",
      "ci_low": "74.63075903413026",
      "ci_high": "82.49180593134915",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "65.34883720930233",
      "ci_low": "58.524497917874655",
      "ci_high": "71.85775568032744",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7317833995639724",
      "ci_low": "0.6337895471807417",
      "ci_high": "0.7515744539043686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7317833995639724",
      "ci_low": "0.6337895471807417",
      "ci_high": "0.7515744539043686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.34444444444444444",
      "ci_low": "0.2827548725637182",
      "ci_high": "0.4079436823104693",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.4444444444444444",
      "ci_low": "0.3787878787878788",
      "ci_high": "0.5104932301740812",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.775",
      "ci_low": "0.6930671356609345",
      "ci_high": "0.8521739130434782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.1",
      "ci_low": "0.0639059898067316",
      "ci_high": "0.14068441064638784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.15925925925925927",
      "ci_low": "0.11244979919678715",
      "ci_high": "0.21011673151750973",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "59.85185185185185",
      "ci_low": "55.468138874849345",
      "ci_high": "64.3097940536103",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "43.925925925925924",
      "ci_low": "36.013355531790786",
      "ci_high": "51.66023166023166",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.3540159640948147",
      "ci_low": "0.30258510974588276",
      "ci_high": "0.38419933529182304",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.41957447596422487",
      "ci_low": "0.3586193893284536",
      "ci_high": "0.45534736034586437",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6484018264840182",
      "ci_low": "0.5810778635778635",
      "ci_high": "0.7123893805309734",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7534246575342466",
      "ci_low": "0.6941747572815534",
      "ci_high": "0.8108108108108109",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.8606060606060606",
      "ci_low": "0.8047337278106509",
      "ci_high": "0.910828025477707",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1050228310502283",
      "ci_low": "0.06666666666666667",
      "ci_high": "0.14754098360655737",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1187214611872146",
      "ci_low": "0.07791955126585477",
      "ci_high": "0.16326783188522928",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "75.57077625570776",
      "ci_low": "70.50239318765774",
      "ci_high": "80.3478369024443",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "63.6986301369863",
      "ci_low": "55.4901093180635",
      "ci_high": "71.25590538748185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6060903608775126",
      "ci_low": "0.5040069871135295",
      "ci_high": "0.6287299489725959",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7346549828818335",
      "ci_low": "0.6109175601376114",
      "ci_high": "0.7620969078455708",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7054794520547946",
      "ci_low": "0.6526278462112729",
      "ci_high": "0.7564950265090714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7876712328767124",
      "ci_low": "0.7399251930501931",
      "ci_high": "0.8327526132404182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8956521739130435",
      "ci_low": "0.8539823008849557",
      "ci_high": "0.933920704845815",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.0821917808219178",
      "ci_low": "0.051901832628476224",
      "ci_high": "0.11552346570397112",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.136986301369863",
      "ci_low": "0.09893353635856425",
      "ci_high": "0.17832167832167833",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "78.39041095890411",
      "ci_low": "74.29420045045046",
      "ci_high": "82.35104207245813",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "64.6917808219178",
      "ci_low": "57.709071969696964",
      "ci_high": "71.39296218487394",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7328920100346514",
      "ci_low": "0.622400462661484",
      "ci_high": "0.7486597525872128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7328920100346514",
      "ci_low": "0.622400462661484",
      "ci_high": "0.7486597525872128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.6631578947368421",
      "ci_low": "0.5867346938775511",
      "ci_high": "0.734375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7526315789473684",
      "ci_low": "0.6827931141439205",
      "ci_high": "0.8195897414230183",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8811188811188811",
      "ci_low": "0.823943661971831",
      "ci_high": "0.93125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.08947368421052632",
      "ci_low": "0.05128205128205128",
      "ci_high": "0.13294940060935673",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.10526315789473684",
      "ci_low": "0.06358381502890173",
      "ci_high": "0.15116420873511058",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "76.68421052631578",
      "ci_low": "70.85106382978724",
      "ci_high": "82.0673433265114",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "66.15789473684211",
      "ci_low": "57.47220077220077",
      "ci_high": "74.26231665228644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5365518836168991",
      "ci_low": "0.4223601772804612",
      "ci_high": "0.5415436765254608",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.6543315653864624",
      "ci_low": "0.5150733869273918",
      "ci_high": "0.6604191177139765",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3595505617977528",
      "ci_low": "0.25316455696202533",
      "ci_high": "0.4672897196261682",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.48314606741573035",
      "ci_low": "0.37142857142857144",
      "ci_high": "0.5918434479054778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.7441860465116279",
      "ci_low": "0.6052631578947368",
      "ci_high": "0.8695652173913043",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.12359550561797752",
      "ci_low": "0.05969705046197584",
      "ci_high": "0.1935483870967742",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.16853932584269662",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.25333333333333335",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "59.10112359550561",
      "ci_low": "51.60458754208754",
      "ci_high": "66.70591822094691",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "42.247191011235955",
      "ci_low": "29.11748366013072",
      "ci_high": "54.77094136418029",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3269601857837152",
      "ci_low": "0.17769841269841266",
      "ci_high": "0.34116606986977355",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.4012693189163778",
      "ci_low": "0.21808441558441555",
      "ci_high": "0.41870381302199483",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "9933",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.23076923076923078",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "61.66666666666667",
      "ci_low": "43.33333333333333",
      "ci_high": "81.53846153846153",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "53.33333333333333",
      "ci_low": "28.571428571428573",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3681592039800995",
      "ci_low": "0.29901960784313725",
      "ci_high": "0.4381443298969072",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.47761194029850745",
      "ci_low": "0.40414507772020725",
      "ci_high": "0.5505105401844531",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.7708333333333334",
      "ci_low": "0.6774193548387096",
      "ci_high": "0.8585858585858586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.10945273631840796",
      "ci_low": "0.06532390211929212",
      "ci_high": "0.15865384615384615",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.16417910447761194",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.22105263157894736",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "61.691542288557216",
      "ci_low": "56.71942767717415",
      "ci_high": "66.56720076090137",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "45.273631840796014",
      "ci_low": "36.0690093488601",
      "ci_high": "54.00052910052909",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.4022299155176415",
      "ci_low": "0.30736279714908743",
      "ci_high": "0.4302338252346421",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.46181953263136616",
      "ci_low": "0.3528980263563597",
      "ci_high": "0.4939721697138482",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.4714285714285714",
      "ci_low": "0.34177215189873417",
      "ci_high": "0.5972267316017316",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.4722222222222222",
      "ci_high": "0.717948717948718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.7857142857142857",
      "ci_low": "0.6521739130434783",
      "ci_high": "0.9069767441860465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.12857142857142856",
      "ci_low": "0.05405405405405406",
      "ci_high": "0.20967741935483872",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.06451612903225806",
      "ci_high": "0.22580645161290322",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "62.285714285714285",
      "ci_low": "52.23880597014926",
      "ci_high": "71.82926829268293",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "48.0",
      "ci_low": "33.6227726300784",
      "ci_high": "61.97592592592592",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.44160202542555477",
      "ci_low": "0.3270558314234785",
      "ci_high": "0.5077939935064935",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6308600363222211",
      "ci_low": "0.46722261631925494",
      "ci_high": "0.7254199907235621",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7111111111111111",
      "ci_low": "0.5686163101604278",
      "ci_high": "0.8461538461538461",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7777777777777778",
      "ci_low": "0.6428571428571429",
      "ci_high": "0.9",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.9142857142857143",
      "ci_low": "0.8055555555555556",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.06666666666666667",
      "ci_low": "0.0",
      "ci_high": "0.15384615384615385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.08888888888888889",
      "ci_low": "0.019230769230769232",
      "ci_high": "0.18518518518518517",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "80.66666666666666",
      "ci_low": "69.75594512195121",
      "ci_high": "90.28571428571428",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "71.77777777777779",
      "ci_low": "54.347084980237156",
      "ci_high": "86.36363636363637",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6798412698412698",
      "ci_low": "0.38388690476190473",
      "ci_high": "0.6933376068376068",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7844322344322344",
      "ci_low": "0.4429464285714286",
      "ci_high": "0.8000049309664694",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3377777777777778",
      "ci_low": "0.26956521739130435",
      "ci_high": "0.4080717488789238",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.4444444444444444",
      "ci_low": "0.37199951690821254",
      "ci_high": "0.5186922135039232",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.76",
      "ci_low": "0.6666666666666666",
      "ci_high": "0.8469387755102041",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.10666666666666667",
      "ci_low": "0.06557198081466104",
      "ci_high": "0.15246826784221323",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.16444444444444445",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.2227494017174229",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "59.15555555555556",
      "ci_low": "54.24951923076923",
      "ci_high": "64.16676008968611",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "42.71111111111111",
      "ci_low": "33.74415339997532",
      "ci_high": "51.30260642491131",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.35609283808296965",
      "ci_low": "0.292805544864707",
      "ci_high": "0.38899839161526467",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.4220359562464826",
      "ci_low": "0.3470287939137268",
      "ci_high": "0.4610351308032766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4788732394366197",
      "ci_low": "0.36065573770491804",
      "ci_high": "0.5974065568577763",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5352112676056338",
      "ci_low": "0.4166666666666667",
      "ci_high": "0.6538580246913579",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8947368421052632",
      "ci_low": "0.7837837837837838",
      "ci_high": "0.9761904761904762",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.056338028169014086",
      "ci_low": "0.01282051282051282",
      "ci_high": "0.11594202898550725",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.09859154929577464",
      "ci_low": "0.034873696872493996",
      "ci_high": "0.17567567567567569",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "68.59154929577466",
      "ci_low": "60.327736647276566",
      "ci_high": "76.61764705882354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "58.732394366197184",
      "ci_low": "44.99837662337662",
      "ci_high": "70.76923076923077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5212522045855379",
      "ci_low": "0.31775132275132273",
      "ci_high": "0.531626017410331",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5212522045855379",
      "ci_low": "0.31775132275132273",
      "ci_high": "0.531626017410331",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.707641196013289",
      "ci_low": "0.6559476886906614",
      "ci_high": "0.7580085524050052",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7873754152823921",
      "ci_low": "0.740484429065744",
      "ci_high": "0.8316831683168316",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8987341772151899",
      "ci_low": "0.8577981651376146",
      "ci_high": "0.9358974358974359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.07973421926910298",
      "ci_low": "0.050292941825199895",
      "ci_high": "0.11224489795918367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.132890365448505",
      "ci_low": "0.09570591018146184",
      "ci_high": "0.17307692307692307",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "78.63787375415282",
      "ci_low": "74.63075903413026",
      "ci_high": "82.49180593134915",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "65.34883720930233",
      "ci_low": "58.524497917874655",
      "ci_high": "71.85775568032744",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7317833995639724",
      "ci_low": "0.6337895471807417",
      "ci_high": "0.7515744539043686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7317833995639724",
      "ci_low": "0.6337895471807417",
      "ci_high": "0.7515744539043686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7692307692307693",
      "ci_low": "0.6626468373493977",
      "ci_high": "0.8641975308641975",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8974358974358975",
      "ci_low": "0.8253869047619047",
      "ci_high": "0.96",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8571428571428571",
      "ci_low": "0.7647058823529411",
      "ci_high": "0.935064935064935",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.1282051282051282",
      "ci_low": "0.057971014492753624",
      "ci_high": "0.20987654320987653",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.14102564102564102",
      "ci_low": "0.06779661016949153",
      "ci_high": "0.22535211267605634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "83.84615384615385",
      "ci_low": "76.3233359133127",
      "ci_high": "90.67590707145501",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "69.74358974358975",
      "ci_low": "54.7824464633355",
      "ci_high": "82.72758405977584",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5129910873440285",
      "ci_low": "0.38842460317460314",
      "ci_high": "0.5395386049622891",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8015485739750445",
      "ci_low": "0.6069134424603175",
      "ci_high": "0.8430290702535769",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7098765432098766",
      "ci_low": "0.6594410313467491",
      "ci_high": "0.7603557385106906",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7870370370370371",
      "ci_low": "0.7405243200354283",
      "ci_high": "0.8317167226204518",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.9019607843137255",
      "ci_low": "0.8634682642007142",
      "ci_high": "0.9367088607594937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.07716049382716049",
      "ci_low": "0.04985314404217027",
      "ci_high": "0.10714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.09567901234567901",
      "ci_low": "0.06432748538011696",
      "ci_high": "0.1298741506171846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "79.22839506172839",
      "ci_low": "75.32667829843803",
      "ci_high": "83.01545731707317",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "69.6604938271605",
      "ci_low": "63.235250088059175",
      "ci_high": "75.64268238069752",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.6462027392780747",
      "ci_low": "0.5791572090196322",
      "ci_high": "0.6688190750504688",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7652400859871937",
      "ci_low": "0.6858440633127222",
      "ci_high": "0.7920225888755551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.4473684210526316",
      "ci_low": "0.32529367469879517",
      "ci_high": "0.5692307692307692",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6447368421052632",
      "ci_low": "0.5238095238095238",
      "ci_high": "0.7600140845070421",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6938775510204082",
      "ci_low": "0.5476078167115903",
      "ci_high": "0.8292682926829268",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.19736842105263158",
      "ci_low": "0.10666666666666667",
      "ci_high": "0.3",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2631578947368421",
      "ci_low": "0.16216216216216217",
      "ci_high": "0.3736607142857138",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "62.10526315789473",
      "ci_low": "52.76849315068493",
      "ci_high": "71.01486572890025",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "35.78947368421053",
      "ci_low": "17.637432795698928",
      "ci_high": "52.64714741318214",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.34795198781266895",
      "ci_low": "0.20830328272351295",
      "ci_high": "0.3856887374321584",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.4175423853752028",
      "ci_low": "0.24996393926821556",
      "ci_high": "0.46282648491859013",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.580952380952381",
      "ci_low": "0.4742268041237113",
      "ci_high": "0.6857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.5634920634920635",
      "ci_high": "0.7678571428571429",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8714285714285714",
      "ci_low": "0.7857142857142857",
      "ci_high": "0.9444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.08571428571428572",
      "ci_low": "0.036036036036036036",
      "ci_high": "0.14285714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.13333333333333333",
      "ci_low": "0.06862745098039216",
      "ci_high": "0.2075565103680178",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "73.90476190476191",
      "ci_low": "67.25472207231516",
      "ci_high": "80.47192489114659",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "60.57142857142857",
      "ci_low": "48.16279192089207",
      "ci_high": "71.76523109243696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5669526975082531",
      "ci_low": "0.4114755691839025",
      "ci_high": "0.5790927515196521",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6378217846967847",
      "ci_low": "0.4629100153318903",
      "ci_high": "0.6514793454596086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.43157894736842106",
      "ci_low": "0.37037037037037035",
      "ci_high": "0.49264705882352944",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.519298245614035",
      "ci_low": "0.4603745470448582",
      "ci_high": "0.5804195804195804",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.831081081081081",
      "ci_low": "0.7664233576642335",
      "ci_high": "0.888199671172817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.08771929824561403",
      "ci_low": "0.05647710348993429",
      "ci_high": "0.12318840579710146",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.1543859649122807",
      "ci_low": "0.11301369863013698",
      "ci_high": "0.19860744494890836",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "63.96491228070175",
      "ci_low": "59.53161694700095",
      "ci_high": "68.36134057971014",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "48.526315789473685",
      "ci_low": "41.44252228932988",
      "ci_high": "55.38194444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.449802005001963",
      "ci_low": "0.37127089388336454",
      "ci_high": "0.47522819991095683",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.47286877448924314",
      "ci_low": "0.3903104269030242",
      "ci_high": "0.49959887682946746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8448979591836735",
      "ci_low": "0.7940074906367042",
      "ci_high": "0.8933345588235294",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9061224489795918",
      "ci_low": "0.8619399081995879",
      "ci_high": "0.9458333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9324324324324325",
      "ci_low": "0.8976715825297787",
      "ci_high": "0.9641255605381166",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.061224489795918366",
      "ci_low": "0.032520325203252036",
      "ci_high": "0.09282700421940929",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.0653061224489796",
      "ci_low": "0.03539562727745966",
      "ci_high": "0.09767616716209125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "91.71428571428571",
      "ci_low": "88.98493184179961",
      "ci_high": "94.28027611940298",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "85.18367346938776",
      "ci_low": "79.72076411960134",
      "ci_high": "90.24693908784818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7884168741311597",
      "ci_low": "0.6433944250002739",
      "ci_high": "0.764112325064706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.841868865597679",
      "ci_low": "0.6870143860172415",
      "ci_high": "0.8159165504928217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6076923076923076",
      "ci_low": "0.5217313664596274",
      "ci_high": "0.695659518213866",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7230769230769231",
      "ci_low": "0.6408395686619719",
      "ci_high": "0.8017260795278036",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.8404255319148937",
      "ci_low": "0.7582417582417582",
      "ci_high": "0.9146341463414634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.11538461538461539",
      "ci_low": "0.06086809635722679",
      "ci_high": "0.1760014084507042",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.14615384615384616",
      "ci_low": "0.08546511627906977",
      "ci_high": "0.21088568239795918",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "77.92307692307692",
      "ci_low": "72.44430894308942",
      "ci_high": "83.42867803837953",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "63.30769230769231",
      "ci_low": "52.99208918978586",
      "ci_high": "73.0701754385965",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5672949735449735",
      "ci_low": "0.4270411532911533",
      "ci_high": "0.5654266826923077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6188672438672438",
      "ci_low": "0.46586307631762175",
      "ci_high": "0.6168291083916084",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.27631578947368424",
      "ci_low": "0.17855919765166342",
      "ci_high": "0.38095238095238093",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.35526315789473684",
      "ci_low": "0.25",
      "ci_high": "0.4657534246575342",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.7777777777777778",
      "ci_low": "0.6071428571428571",
      "ci_high": "0.9259259259259259",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.07894736842105263",
      "ci_low": "0.025",
      "ci_high": "0.14516129032258066",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.14473684210526316",
      "ci_low": "0.07042253521126761",
      "ci_high": "0.22857142857142856",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "41.578947368421055",
      "ci_low": "32.857142857142854",
      "ci_high": "50.588235294117645",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "27.10526315789474",
      "ci_low": "13.103243021346469",
      "ci_high": "40.71450086058519",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.35763888888888884",
      "ci_low": "0.1581345390720391",
      "ci_high": "0.3180654761904761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3731884057971015",
      "ci_low": "0.1650099538143016",
      "ci_high": "0.33189440993788805",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.13541666666666666",
      "ci_high": "0.3023332313341492",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.32653061224489793",
      "ci_low": "0.23333333333333334",
      "ci_high": "0.4226903251387786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.65625",
      "ci_low": "0.4827586206896552",
      "ci_high": "0.8181818181818182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.11224489795918367",
      "ci_low": "0.05319148936170213",
      "ci_high": "0.1794943820224718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.20408163265306123",
      "ci_low": "0.125",
      "ci_high": "0.28846153846153844",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "40.91836734693878",
      "ci_low": "33.53655457009115",
      "ci_high": "48.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "20.51020408163265",
      "ci_low": "8.105131578947368",
      "ci_high": "32.88659793814433",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.19630192502532925",
      "ci_low": "0.10437310030395137",
      "ci_high": "0.2128021092382794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.20056935817805382",
      "ci_low": "0.10664208074534162",
      "ci_high": "0.2174282420478072",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5794392523364486",
      "ci_low": "0.48484848484848486",
      "ci_high": "0.6704614745011085",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7009345794392523",
      "ci_low": "0.6131999391357273",
      "ci_high": "0.7840909090909091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.8266666666666667",
      "ci_low": "0.734375",
      "ci_high": "0.9090909090909091",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.12149532710280374",
      "ci_low": "0.0625",
      "ci_high": "0.18811881188118812",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.16822429906542055",
      "ci_low": "0.09900990099009901",
      "ci_high": "0.24347826086956523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "71.96261682242991",
      "ci_low": "65.04664770418528",
      "ci_high": "78.54576048951049",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "55.14018691588785",
      "ci_low": "42.352791116446575",
      "ci_high": "67.2340857983048",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5518796992481203",
      "ci_low": "0.3559276663881927",
      "ci_high": "0.5115302144249513",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6168067226890757",
      "ci_low": "0.3978015094926859",
      "ci_high": "0.5717102396514161",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.42857142857142855",
      "ci_low": "0.34285714285714286",
      "ci_high": "0.514290890269151",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5338345864661654",
      "ci_low": "0.44881166049096805",
      "ci_high": "0.6201550387596899",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.8028169014084507",
      "ci_low": "0.704225352112676",
      "ci_high": "0.8902480788506515",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.10526315789473684",
      "ci_low": "0.05714285714285714",
      "ci_high": "0.16058394160583941",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.18796992481203006",
      "ci_low": "0.12295081967213115",
      "ci_high": "0.25925925925925924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "66.24060150375941",
      "ci_low": "60.4165805785124",
      "ci_high": "71.90476190476191",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "47.443609022556394",
      "ci_low": "36.51477272727273",
      "ci_high": "57.43141284403669",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.4678095238095238",
      "ci_low": "0.3112497308247308",
      "ci_high": "0.4589404952890247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.47735665694849366",
      "ci_low": "0.31760176614768454",
      "ci_high": "0.46830662784594357",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / MEDIUM thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6876456876456877",
      "ci_low": "0.6402873628698851",
      "ci_high": "0.7333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7645687645687645",
      "ci_low": "0.7213086424416408",
      "ci_high": "0.8069309615343313",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.899390243902439",
      "ci_low": "0.8660428501618715",
      "ci_high": "0.9300911854103343",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.07692307692307693",
      "ci_low": "0.052995391705069124",
      "ci_high": "0.10267884430667132",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.10023310023310024",
      "ci_low": "0.07242990654205607",
      "ci_high": "0.13013908989936385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "78.32167832167832",
      "ci_low": "74.93382958018756",
      "ci_high": "81.62621759308652",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "68.2983682983683",
      "ci_low": "62.650375085246644",
      "ci_high": "73.5150043888523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5889353821931136",
      "ci_low": "0.5126891394994124",
      "ci_high": "0.5943557395231674",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7314197488527379",
      "ci_low": "0.6367268345395928",
      "ci_high": "0.7381514829561918",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4293628808864266",
      "ci_low": "0.3771917293233083",
      "ci_high": "0.48265915221579964",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5512465373961218",
      "ci_low": "0.49865210848873986",
      "ci_high": "0.6034493507610285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.7788944723618091",
      "ci_low": "0.7184452070081464",
      "ci_high": "0.835",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.12188365650969529",
      "ci_low": "0.08963376395334254",
      "ci_high": "0.15691764111130643",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.1883656509695291",
      "ci_low": "0.14763002876843692",
      "ci_high": "0.2305393379906853",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "63.18559556786703",
      "ci_low": "59.280031909487654",
      "ci_high": "67.14681440443213",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "44.34903047091413",
      "ci_low": "37.581457379862705",
      "ci_high": "50.814677540336156",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4066446751628938",
      "ci_low": "0.32916461914704476",
      "ci_high": "0.4191355405619619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4479983709421711",
      "ci_low": "0.3626389871958967",
      "ci_high": "0.46175949383944964",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.768",
      "ci_low": "0.7222222222222222",
      "ci_high": "0.8128491620111732",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8426666666666667",
      "ci_low": "0.8016991232373748",
      "ci_high": "0.8810810810810811",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.9113924050632911",
      "ci_low": "0.8786861677054417",
      "ci_high": "0.9409838611139395",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.07466666666666667",
      "ci_low": "0.049479166666666664",
      "ci_high": "0.10236220472440945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.09066666666666667",
      "ci_low": "0.06284117243155093",
      "ci_high": "0.12096774193548387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "87.33333333333333",
      "ci_low": "84.66141402159687",
      "ci_high": "89.88637968669747",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "78.26666666666667",
      "ci_low": "73.21612169163983",
      "ci_high": "83.05043220896363",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7129652754652754",
      "ci_low": "0.6376155725448074",
      "ci_high": "0.7178825557336174",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7678087581933735",
      "ci_low": "0.6866629242790233",
      "ci_high": "0.7731042907900493",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.20689655172413793",
      "ci_low": "0.14917127071823205",
      "ci_high": "0.26993865030674846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3563218390804598",
      "ci_low": "0.2857142857142857",
      "ci_high": "0.4301675977653631",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5806451612903226",
      "ci_low": "0.45651220442410373",
      "ci_high": "0.704225352112676",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.14942528735632185",
      "ci_low": "0.09815950920245399",
      "ci_high": "0.2057142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.23563218390804597",
      "ci_low": "0.1728395061728395",
      "ci_high": "0.30120481927710846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "38.56321839080459",
      "ci_low": "33.166522988505754",
      "ci_high": "44.17660427807487",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "15.0",
      "ci_low": "5.6279697303083465",
      "ci_high": "24.327763854079638",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.24082892416225748",
      "ci_low": "0.13113870697204033",
      "ci_high": "0.22645930201485756",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.2533395176252319",
      "ci_low": "0.13795110733422425",
      "ci_high": "0.23822342160004498",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5208333333333334",
      "ci_low": "0.4570106490872211",
      "ci_high": "0.5833333333333334",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6166666666666667",
      "ci_low": "0.5542168674698795",
      "ci_high": "0.6785714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.8445945945945946",
      "ci_low": "0.7832167832167832",
      "ci_high": "0.9",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.09583333333333334",
      "ci_low": "0.060603005865102644",
      "ci_high": "0.13524821274497717",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.15",
      "ci_low": "0.10384615384615385",
      "ci_high": "0.2",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "70.20833333333333",
      "ci_low": "65.73142310872261",
      "ci_high": "74.5455476900149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "55.20833333333333",
      "ci_low": "46.992457308525545",
      "ci_high": "62.71514825442372",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5414579606440071",
      "ci_low": "0.40863227857413903",
      "ci_high": "0.5186205636978279",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5894352482960077",
      "ci_low": "0.44484020199210067",
      "ci_high": "0.5645742845318128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.66",
      "ci_low": "0.6120481927710844",
      "ci_high": "0.707808564231738",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.765",
      "ci_low": "0.7220750394465348",
      "ci_high": "0.8070175438596491",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.8627450980392157",
      "ci_low": "0.8228218725054168",
      "ci_high": "0.8993294038176589",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.105",
      "ci_low": "0.07673267326732673",
      "ci_high": "0.1358974358974359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1275",
      "ci_low": "0.09660574412532637",
      "ci_high": "0.1605287019361876",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "75.65",
      "ci_low": "72.08876482701548",
      "ci_high": "79.16467035217035",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "62.9",
      "ci_low": "56.80097026209677",
      "ci_high": "68.82352941176471",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5410575647262954",
      "ci_low": "0.5022617742238907",
      "ci_high": "0.5616383998065115",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7304277123804986",
      "ci_low": "0.6780533952022526",
      "ci_high": "0.7582118397387907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.47692307692307695",
      "ci_low": "0.4239394719722433",
      "ci_high": "0.5298512308586936",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5666666666666667",
      "ci_low": "0.5149981343283582",
      "ci_high": "0.618798955613577",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.8416289592760181",
      "ci_low": "0.7912621359223301",
      "ci_high": "0.8879310344827587",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.08974358974358974",
      "ci_low": "0.0625",
      "ci_high": "0.11928934010152284",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.15384615384615385",
      "ci_low": "0.11764705882352941",
      "ci_high": "0.1935503636938646",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "67.05128205128206",
      "ci_low": "63.30684188125133",
      "ci_high": "70.809450887043",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "51.66666666666667",
      "ci_low": "45.23179639858897",
      "ci_high": "57.82614185331576",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5087793128538374",
      "ci_low": "0.4570642681703775",
      "ci_high": "0.536713242085056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5479161830733634",
      "ci_low": "0.4922230580296374",
      "ci_high": "0.5779988760915987",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6943521594684385",
      "ci_low": "0.6411939875212705",
      "ci_high": "0.7442641519444002",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7973421926910299",
      "ci_low": "0.7516339869281046",
      "ci_high": "0.8412162162162162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8708333333333333",
      "ci_low": "0.8262665806927045",
      "ci_high": "0.9108527131782945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.10299003322259136",
      "ci_low": "0.07028620428050904",
      "ci_high": "0.13915857605177995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.15282392026578073",
      "ci_low": "0.11262798634812286",
      "ci_high": "0.19472039900619273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "77.77408637873754",
      "ci_low": "73.745925615243",
      "ci_high": "81.6447474533107",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "62.49169435215947",
      "ci_low": "55.3740402913577",
      "ci_high": "69.4185606060606",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7004478170943473",
      "ci_low": "0.6040141262269906",
      "ci_high": "0.7192537655538894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7179590125217061",
      "ci_low": "0.6191144793826655",
      "ci_high": "0.7372351096927368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.362962962962963",
      "ci_low": "0.29963828681397403",
      "ci_high": "0.4271261618370693",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.45185185185185184",
      "ci_low": "0.3854527333131984",
      "ci_high": "0.517485595390525",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.8032786885245902",
      "ci_low": "0.7297297297297297",
      "ci_high": "0.8714285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.08888888888888889",
      "ci_low": "0.05645009701673539",
      "ci_high": "0.12400077519379844",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.14814814814814814",
      "ci_low": "0.10416666666666667",
      "ci_high": "0.19494584837545126",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "61.25925925925926",
      "ci_low": "56.78690216326645",
      "ci_high": "65.80016129032258",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "46.44444444444445",
      "ci_low": "38.78463821684588",
      "ci_high": "53.85249305593377",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.34903812218053704",
      "ci_low": "0.3011136493100275",
      "ci_high": "0.3805566195229592",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.43952948719030593",
      "ci_low": "0.37918015098299757",
      "ci_high": "0.4792194468066894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6529680365296804",
      "ci_low": "0.5857142857142857",
      "ci_high": "0.7157451819667899",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7534246575342466",
      "ci_low": "0.6952351916376306",
      "ci_high": "0.8086124401913876",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.8666666666666667",
      "ci_low": "0.8117627960275019",
      "ci_high": "0.9162127560521414",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1004566210045662",
      "ci_low": "0.06302438298153908",
      "ci_high": "0.14220183486238533",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1141552511415525",
      "ci_low": "0.07373271889400922",
      "ci_high": "0.1588785046728972",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "75.15981735159818",
      "ci_low": "70.0",
      "ci_high": "79.9613095238095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "63.74429223744293",
      "ci_low": "55.412844036697244",
      "ci_high": "71.3397849105086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5377082820000777",
      "ci_low": "0.4585190219469526",
      "ci_high": "0.5605547246314292",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7169443760001035",
      "ci_low": "0.6113586959292701",
      "ci_high": "0.7474062995085723",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6917808219178082",
      "ci_low": "0.6378689392845105",
      "ci_high": "0.7429601917057903",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.797945205479452",
      "ci_low": "0.7516329461720994",
      "ci_high": "0.8424657534246576",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8669527896995708",
      "ci_low": "0.8217391304347826",
      "ci_high": "0.9078947368421053",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.10616438356164383",
      "ci_low": "0.0727256999056307",
      "ci_high": "0.14336950506008855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.15753424657534246",
      "ci_low": "0.11660520086452365",
      "ci_high": "0.20068056634418474",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "77.5",
      "ci_low": "73.36952747584542",
      "ci_high": "81.50544685351666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "61.74657534246576",
      "ci_low": "54.489683373349344",
      "ci_high": "68.85417472910217",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.701649912612982",
      "ci_low": "0.5939673081964599",
      "ci_high": "0.7161772561477738",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7187633251157377",
      "ci_low": "0.6084543157134467",
      "ci_high": "0.7336449941025975",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.6684210526315789",
      "ci_low": "0.593403964456596",
      "ci_high": "0.7393617021276596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7526315789473684",
      "ci_low": "0.6834845097044857",
      "ci_high": "0.8186904761904761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8881118881118881",
      "ci_low": "0.8333333333333334",
      "ci_high": "0.9375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.08421052631578947",
      "ci_low": "0.04705606186136427",
      "ci_high": "0.12568306010928962",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.10526315789473684",
      "ci_low": "0.06341463414634146",
      "ci_high": "0.15217667181637604",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "76.26315789473685",
      "ci_low": "70.36827788382875",
      "ci_high": "81.7858082706767",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "65.73684210526315",
      "ci_low": "56.85554906874751",
      "ci_high": "73.94455128205128",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.5153830905870082",
      "ci_low": "0.4119558066524979",
      "ci_high": "0.5221358204686132",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.6536566026957177",
      "ci_low": "0.5224805352665826",
      "ci_high": "0.6622210405943387",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.3707865168539326",
      "ci_low": "0.26373626373626374",
      "ci_high": "0.4803921568627451",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.48314606741573035",
      "ci_low": "0.3655913978494624",
      "ci_high": "0.5957446808510638",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.7674418604651163",
      "ci_low": "0.6388888888888888",
      "ci_high": "0.8867924528301887",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.11235955056179775",
      "ci_low": "0.05127551020408164",
      "ci_high": "0.1797752808988764",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.15730337078651685",
      "ci_low": "0.08139534883720931",
      "ci_high": "0.24050632911392406",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "58.426966292134836",
      "ci_low": "50.28558077436582",
      "ci_high": "66.63448844884488",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "42.69662921348314",
      "ci_low": "29.814444444444444",
      "ci_high": "55.24456163480553",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.31993661100803955",
      "ci_low": "0.17556755196557827",
      "ci_high": "0.3348700787133035",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.4071920503738685",
      "ci_low": "0.2234496115925542",
      "ci_high": "0.4261982819987498",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "9933",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.08333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.23076923076923078",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "63.33333333333333",
      "ci_low": "45.0",
      "ci_high": "82.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "30.0",
      "ci_high": "81.81818181818181",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4222222222222222",
      "ci_low": "0.08888888888888889",
      "ci_high": "0.4444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3880597014925373",
      "ci_low": "0.31817658148303307",
      "ci_high": "0.45959895160132125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.48756218905472637",
      "ci_low": "0.4162673353896971",
      "ci_high": "0.5588235294117647",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.7959183673469388",
      "ci_low": "0.7124802927927928",
      "ci_high": "0.8738738738738738",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.09950248756218906",
      "ci_low": "0.06",
      "ci_high": "0.14218009478672985",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.14925373134328357",
      "ci_low": "0.1004665071770335",
      "ci_high": "0.20183706774876498",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "63.73134328358209",
      "ci_low": "58.87634122886932",
      "ci_high": "68.61114433811802",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "48.805970149253724",
      "ci_low": "40.310840767053044",
      "ci_high": "57.15861267433324",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.3959636323272687",
      "ci_low": "0.31056102070124975",
      "ci_high": "0.424587694642541",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.4839555506222173",
      "ci_low": "0.37957458085708295",
      "ci_high": "0.5189405156742167",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "6374",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.4714285714285714",
      "ci_low": "0.3424657534246575",
      "ci_high": "0.5970201077943614",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6",
      "ci_low": "0.4714285714285714",
      "ci_high": "0.7166666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.7857142857142857",
      "ci_low": "0.6521739130434783",
      "ci_high": "0.9047619047619048",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.12857142857142856",
      "ci_low": "0.05555555555555555",
      "ci_high": "0.20968239745121459",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.06666666666666667",
      "ci_high": "0.22807017543859648",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "61.28571428571429",
      "ci_low": "50.920753588516746",
      "ci_high": "71.14754098360656",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "47.0",
      "ci_low": "31.714191729323307",
      "ci_high": "61.60017857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3692226890756302",
      "ci_low": "0.27354372683416806",
      "ci_high": "0.4243589917027416",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6329531812725089",
      "ci_low": "0.46893210314428807",
      "ci_high": "0.7274725572047002",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6888888888888889",
      "ci_low": "0.5434782608695652",
      "ci_high": "0.8205128205128205",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7777777777777778",
      "ci_low": "0.6470588235294118",
      "ci_high": "0.8928571428571429",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8857142857142857",
      "ci_low": "0.7631578947368421",
      "ci_high": "0.975",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.08888888888888889",
      "ci_low": "0.0196078431372549",
      "ci_high": "0.18604651162790697",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.1111111111111111",
      "ci_low": "0.025",
      "ci_high": "0.21621621621621623",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "77.11111111111111",
      "ci_low": "65.71428571428571",
      "ci_high": "87.25056818181817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "66.0",
      "ci_low": "47.0",
      "ci_high": "82.32626538987688",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5462184873949579",
      "ci_low": "0.3078417366946779",
      "ci_high": "0.5659371021135726",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7142857142857142",
      "ci_low": "0.40256227106227105",
      "ci_high": "0.740071595071595",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.3466666666666667",
      "ci_low": "0.27755102040816326",
      "ci_high": "0.41810344827586204",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.44",
      "ci_low": "0.3673469387755102",
      "ci_high": "0.5130446164352379",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.7878787878787878",
      "ci_low": "0.7040816326530612",
      "ci_high": "0.8653886554621848",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.09333333333333334",
      "ci_low": "0.05752085240565558",
      "ci_high": "0.13215859030837004",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.1511111111111111",
      "ci_low": "0.10222222222222223",
      "ci_high": "0.203125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "60.08888888888889",
      "ci_low": "55.1281512605042",
      "ci_high": "65.15837104072398",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "44.977777777777774",
      "ci_low": "36.59367344765151",
      "ci_high": "53.23156788490459",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.33647257025739996",
      "ci_low": "0.2750505514305566",
      "ci_high": "0.36945036951581256",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.42370619958339256",
      "ci_low": "0.34635995365329353",
      "ci_high": "0.4652337986495418",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5211267605633803",
      "ci_low": "0.40625",
      "ci_high": "0.6363965744400523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5915492957746479",
      "ci_low": "0.47692307692307695",
      "ci_high": "0.7049180327868853",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8809523809523809",
      "ci_low": "0.772707336523126",
      "ci_high": "0.9736842105263158",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.07042253521126761",
      "ci_low": "0.014925373134328358",
      "ci_high": "0.13513513513513514",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.11267605633802817",
      "ci_low": "0.045454545454545456",
      "ci_high": "0.1917808219178082",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "70.98591549295774",
      "ci_low": "62.83292253521126",
      "ci_high": "78.84615384615385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "59.71830985915493",
      "ci_low": "46.44736842105263",
      "ci_high": "71.80555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5494708994708994",
      "ci_low": "0.35202776647221096",
      "ci_high": "0.5645342312008979",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5494708994708994",
      "ci_low": "0.35202776647221096",
      "ci_high": "0.5645342312008979",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6943521594684385",
      "ci_low": "0.6411939875212705",
      "ci_high": "0.7442641519444002",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7973421926910299",
      "ci_low": "0.7516339869281046",
      "ci_high": "0.8412162162162162",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8708333333333333",
      "ci_low": "0.8262665806927045",
      "ci_high": "0.9108527131782945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.10299003322259136",
      "ci_low": "0.07028620428050904",
      "ci_high": "0.13915857605177995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.15282392026578073",
      "ci_low": "0.11262798634812286",
      "ci_high": "0.19472039900619273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "77.77408637873754",
      "ci_low": "73.745925615243",
      "ci_high": "81.6447474533107",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "62.49169435215947",
      "ci_low": "55.3740402913577",
      "ci_high": "69.4185606060606",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7004478170943473",
      "ci_low": "0.6040141262269906",
      "ci_high": "0.7192537655538894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7179590125217061",
      "ci_low": "0.6191144793826655",
      "ci_high": "0.7372351096927368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7948717948717948",
      "ci_low": "0.6984126984126984",
      "ci_high": "0.8813559322033898",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8846153846153846",
      "ci_low": "0.8089887640449438",
      "ci_high": "0.9506172839506173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8985507246376812",
      "ci_low": "0.8169014084507042",
      "ci_high": "0.9655172413793104",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.08974358974358974",
      "ci_low": "0.030303030303030304",
      "ci_high": "0.16129032258064516",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.10256410256410256",
      "ci_low": "0.04",
      "ci_high": "0.17857142857142858",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "85.64102564102564",
      "ci_low": "78.50727611940299",
      "ci_high": "91.8293112332532",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "75.38461538461539",
      "ci_low": "61.72400323275862",
      "ci_high": "86.71647621268656",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5767902813299234",
      "ci_low": "0.44619050663743753",
      "ci_high": "0.6023286176738862",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8291360294117647",
      "ci_low": "0.6413988532913165",
      "ci_high": "0.8658473879062114",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7129629629629629",
      "ci_low": "0.6615372670807453",
      "ci_high": "0.7628578792341679",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7870370370370371",
      "ci_low": "0.7412082962342219",
      "ci_high": "0.8317460317460318",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.9058823529411765",
      "ci_low": "0.8686411135120535",
      "ci_high": "0.9395161290322581",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.07407407407407407",
      "ci_low": "0.04745461839334069",
      "ci_high": "0.10377440099236142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.09259259259259259",
      "ci_low": "0.0626956908248725",
      "ci_high": "0.12538226299694188",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "79.22839506172839",
      "ci_low": "75.33122767250926",
      "ci_high": "83.0149368541906",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "69.96913580246914",
      "ci_low": "63.616785618098575",
      "ci_high": "75.88068574808882",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5983094983387549",
      "ci_low": "0.5386304688811944",
      "ci_high": "0.6184960786156588",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.771504353121026",
      "ci_low": "0.6945498151362771",
      "ci_high": "0.7975344171622969",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.4342105263157895",
      "ci_low": "0.313953488372093",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6710526315789473",
      "ci_low": "0.5569620253164557",
      "ci_high": "0.7812589798850573",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6470588235294118",
      "ci_low": "0.5068661971830988",
      "ci_high": "0.7777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14544721407624633",
      "ci_high": "0.3384615384615385",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.27631578947368424",
      "ci_low": "0.1794871794871795",
      "ci_high": "0.3823620188816266",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "60.39473684210527",
      "ci_low": "51.15928093645485",
      "ci_high": "69.33376623376624",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "32.76315789473684",
      "ci_low": "15.428399311531843",
      "ci_high": "49.37538109756097",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3142857142857143",
      "ci_low": "0.19118086797008363",
      "ci_high": "0.34836276120998344",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.42428571428571427",
      "ci_low": "0.2580941717596129",
      "ci_high": "0.4702897276334776",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6095238095238096",
      "ci_low": "0.5",
      "ci_high": "0.7142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6952380952380952",
      "ci_low": "0.5899935897435897",
      "ci_high": "0.7962962962962963",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8767123287671232",
      "ci_low": "0.796875",
      "ci_high": "0.948051948051948",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.08571428571428572",
      "ci_low": "0.036036036036036036",
      "ci_high": "0.14285714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.12380952380952381",
      "ci_low": "0.06",
      "ci_high": "0.19642857142857142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "75.52380952380952",
      "ci_low": "68.57142857142857",
      "ci_high": "82.40009615384616",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "63.14285714285714",
      "ci_low": "50.792014069828035",
      "ci_high": "74.51026637069923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.5725306789382419",
      "ci_low": "0.4168189666063947",
      "ci_high": "0.5925458024118739",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.6679524587612823",
      "ci_low": "0.48628879437412725",
      "ci_high": "0.6913034361471861",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4280701754385965",
      "ci_low": "0.36899845594200625",
      "ci_high": "0.48829625884732053",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.519298245614035",
      "ci_low": "0.458955223880597",
      "ci_high": "0.5787545787545788",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.8243243243243243",
      "ci_low": "0.7597390013087687",
      "ci_high": "0.8843602825745682",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.0912280701754386",
      "ci_low": "0.058823529411764705",
      "ci_high": "0.1263951202837243",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.1649122807017544",
      "ci_low": "0.12177121771217712",
      "ci_high": "0.211222284200251",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "63.929824561403514",
      "ci_low": "59.47703818369453",
      "ci_high": "68.35582847397953",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "47.43859649122807",
      "ci_low": "40.31463068181819",
      "ci_high": "54.54545454545454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.45035712094922625",
      "ci_low": "0.3745020999083098",
      "ci_high": "0.4754769636778676",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.4619047394351038",
      "ci_low": "0.3841047178546768",
      "ci_high": "0.48766868069524877",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8448979591836735",
      "ci_low": "0.7934262600395763",
      "ci_high": "0.8931623931623932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8938775510204081",
      "ci_low": "0.847925829230257",
      "ci_high": "0.9354838709677419",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9452054794520548",
      "ci_low": "0.91324200913242",
      "ci_high": "0.9734513274336283",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.04897959183673469",
      "ci_low": "0.023622047244094488",
      "ci_high": "0.07751937984496124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.05714285714285714",
      "ci_low": "0.029661016949152543",
      "ci_high": "0.08823529411764706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "91.63265306122449",
      "ci_low": "88.83606785317019",
      "ci_high": "94.252387686506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "85.91836734693878",
      "ci_low": "80.50175779406675",
      "ci_high": "90.70834153543306",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7860630551106741",
      "ci_low": "0.6448308171493854",
      "ci_high": "0.765472030563067",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8393554656266521",
      "ci_low": "0.6885481606849368",
      "ci_high": "0.8173684394148003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6230769230769231",
      "ci_low": "0.5352112676056338",
      "ci_high": "0.7096835754740212",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7461538461538462",
      "ci_low": "0.6666666666666666",
      "ci_high": "0.8222222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.8350515463917526",
      "ci_low": "0.7582330827067669",
      "ci_high": "0.9052667140825036",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.12307692307692308",
      "ci_low": "0.06956521739130435",
      "ci_high": "0.1810344827586207",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.15384615384615385",
      "ci_low": "0.0948905109489051",
      "ci_high": "0.21768844634628043",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "79.23076923076923",
      "ci_low": "73.81652329953288",
      "ci_high": "84.34125379170881",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "63.846153846153854",
      "ci_low": "53.58182302771855",
      "ci_high": "73.45895731063196",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5785545423843296",
      "ci_low": "0.4417746231310061",
      "ci_high": "0.58267972275951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.618001443001443",
      "ci_low": "0.4718956201626657",
      "ci_high": "0.6224078856749311",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.23684210526315788",
      "ci_low": "0.14492753623188406",
      "ci_high": "0.33707865168539325",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3815789473684211",
      "ci_low": "0.2714285714285714",
      "ci_high": "0.49230769230769234",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.6206896551724138",
      "ci_low": "0.43586956521739134",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.14473684210526316",
      "ci_low": "0.07042253521126761",
      "ci_high": "0.22917520491803264",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.19736842105263158",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.2916666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "39.21052631578947",
      "ci_low": "30.920753588516746",
      "ci_high": "47.89473684210527",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "19.473684210526315",
      "ci_low": "5.121878629500581",
      "ci_high": "33.334269662921336",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.29084967320261434",
      "ci_low": "0.1222222222222222",
      "ci_high": "0.2619047619047619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.322463768115942",
      "ci_low": "0.1355072463768116",
      "ci_high": "0.2903726708074534",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.1836734693877551",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.26732673267326734",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.336734693877551",
      "ci_low": "0.24489795918367346",
      "ci_high": "0.4380952380952381",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.5454545454545454",
      "ci_low": "0.37140211640211646",
      "ci_high": "0.7142857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.15306122448979592",
      "ci_low": "0.08571428571428572",
      "ci_high": "0.23",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2653061224489796",
      "ci_low": "0.17755792445482865",
      "ci_high": "0.3592233009708738",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "38.06122448979592",
      "ci_low": "30.988839285714285",
      "ci_high": "45.6945564516129",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "11.53061224489796",
      "ci_low": "-1.1235955056179776",
      "ci_high": "24.28609831029185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.15170068027210884",
      "ci_low": "0.08027210884353742",
      "ci_high": "0.17067077480342782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.1615942028985507",
      "ci_low": "0.0855072463768116",
      "ci_high": "0.18180147750799922",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6448598130841121",
      "ci_low": "0.5517241379310345",
      "ci_high": "0.7327586206896551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7383177570093458",
      "ci_low": "0.6526315789473685",
      "ci_high": "0.8198243243243242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.8734177215189873",
      "ci_low": "0.7948717948717948",
      "ci_high": "0.9431818181818182",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.09345794392523364",
      "ci_low": "0.041237113402061855",
      "ci_high": "0.15315315315315314",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.1308411214953271",
      "ci_low": "0.0707070707070707",
      "ci_high": "0.2",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "75.42056074766356",
      "ci_low": "68.43111455108358",
      "ci_high": "81.75925925925925",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "62.336448598130836",
      "ci_low": "50.21978021978022",
      "ci_high": "73.18619469026548",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5935139573070608",
      "ci_low": "0.3889114668856049",
      "ci_high": "0.549341133004926",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6749766573295986",
      "ci_low": "0.4422914721444134",
      "ci_high": "0.6247408963585432",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.42105263157894735",
      "ci_low": "0.33793103448275863",
      "ci_high": "0.5038167938931297",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.518796992481203",
      "ci_low": "0.4338235294117647",
      "ci_high": "0.6033057851239669",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.8115942028985508",
      "ci_low": "0.7160493827160493",
      "ci_high": "0.8983050847457628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.09774436090225563",
      "ci_low": "0.05042016806722689",
      "ci_high": "0.15151515151515152",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.16541353383458646",
      "ci_low": "0.10317460317460317",
      "ci_high": "0.23484975594874924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "66.01503759398496",
      "ci_low": "60.25194586031134",
      "ci_high": "71.5873015873016",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "49.473684210526315",
      "ci_low": "39.199629629629634",
      "ci_high": "59.00017730496454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.4639014047177312",
      "ci_low": "0.3063629737609329",
      "ci_high": "0.4534023980405559",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.4639014047177312",
      "ci_low": "0.3063629737609329",
      "ci_high": "0.4534023980405559",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / LOW thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7319347319347319",
      "ci_low": "0.6877931616850425",
      "ci_high": "0.7739745180962888",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8531468531468531",
      "ci_low": "0.8173076923076923",
      "ci_high": "0.8868384938549561",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8579234972677595",
      "ci_low": "0.8205128205128205",
      "ci_high": "0.8931512946628186",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.12121212121212122",
      "ci_low": "0.09070294784580499",
      "ci_high": "0.15328510191433686",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.13752913752913754",
      "ci_low": "0.1047613110081938",
      "ci_high": "0.1721870147976805",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "81.86480186480188",
      "ci_low": "78.75782033743731",
      "ci_high": "84.82504331683168",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "68.1118881118881",
      "ci_low": "62.17104883311049",
      "ci_high": "73.68301435406698",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.6111803656842256",
      "ci_low": "0.5328345661557514",
      "ci_high": "0.6144362888196777",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.7787620788557069",
      "ci_low": "0.6789343665532962",
      "ci_high": "0.7829107551089441",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5318559556786704",
      "ci_low": "0.47780523897699717",
      "ci_high": "0.5849596220689877",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.6786703601108033",
      "ci_low": "0.6287243104526535",
      "ci_high": "0.7277936962750716",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.7836734693877551",
      "ci_low": "0.7288135593220338",
      "ci_high": "0.8347835791925465",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.14681440443213298",
      "ci_low": "0.11080332409972299",
      "ci_high": "0.18571623731459794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.21329639889196675",
      "ci_low": "0.17109047508150907",
      "ci_high": "0.25806451612903225",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "69.30747922437673",
      "ci_low": "65.32392199349945",
      "ci_high": "73.23455986664776",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "47.97783933518006",
      "ci_low": "40.79448756257519",
      "ci_high": "55.04400460418359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.4679993612043887",
      "ci_low": "0.38971527884421053",
      "ci_high": "0.47986115462039364",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5314569017066787",
      "ci_low": "0.44255802851800186",
      "ci_high": "0.5449270738909553",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8133333333333334",
      "ci_low": "0.7716535433070866",
      "ci_high": "0.8532345068137572",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.9066666666666666",
      "ci_low": "0.8743311855293534",
      "ci_high": "0.9375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8970588235294118",
      "ci_low": "0.8643067846607669",
      "ci_high": "0.927377255312326",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.09333333333333334",
      "ci_low": "0.06557377049180328",
      "ci_high": "0.12335958005249344",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.112",
      "ci_low": "0.082010582010582",
      "ci_high": "0.1447730335926398",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "89.76",
      "ci_low": "87.38504641909815",
      "ci_high": "92.06815683665499",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "78.56",
      "ci_low": "73.43282573618394",
      "ci_high": "83.38462532299742",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.7348382989559459",
      "ci_low": "0.6601597557648675",
      "ci_high": "0.7394745897377476",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.8007853257853257",
      "ci_low": "0.7194048620514581",
      "ci_high": "0.8058376939449814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3103448275862069",
      "ci_low": "0.2422360248447205",
      "ci_high": "0.3801232375417069",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5287356321839081",
      "ci_low": "0.45454545454545453",
      "ci_high": "0.603448275862069",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5869565217391305",
      "ci_low": "0.4842105263157895",
      "ci_high": "0.6888888888888889",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.21839080459770116",
      "ci_low": "0.15789473684210525",
      "ci_high": "0.28160919540229884",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.28735632183908044",
      "ci_low": "0.22042782941498087",
      "ci_high": "0.35542168674698793",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "47.35632183908046",
      "ci_low": "41.20565239682133",
      "ci_high": "53.410404624277454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "18.620689655172413",
      "ci_low": "8.16970926301555",
      "ci_high": "29.085972558514925",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3339663643235071",
      "ci_low": "0.19873754950698228",
      "ci_high": "0.30663355989248847",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.3643269428983714",
      "ci_low": "0.21680459946216246",
      "ci_high": "0.3345093380645328",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6083333333333333",
      "ci_low": "0.5439983263598327",
      "ci_high": "0.6707818930041153",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.7416666666666667",
      "ci_low": "0.6848739495798319",
      "ci_high": "0.7974151571366216",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.8202247191011236",
      "ci_low": "0.7597765363128491",
      "ci_high": "0.8757070504462943",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.13333333333333333",
      "ci_low": "0.09130434782608696",
      "ci_high": "0.1799163179916318",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.18333333333333332",
      "ci_low": "0.13274336283185842",
      "ci_high": "0.2384937238493724",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "75.58333333333334",
      "ci_low": "71.27189387402933",
      "ci_high": "79.83124135730242",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "57.25",
      "ci_low": "48.67255719164249",
      "ci_high": "65.36173215511324",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.604353086249638",
      "ci_low": "0.4608287727316632",
      "ci_high": "0.5789137339352856",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.6655533987812469",
      "ci_low": "0.5074949775652493",
      "ci_high": "0.637537909523669",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7",
      "ci_low": "0.6542546575632476",
      "ci_high": "0.744899247577819",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.8475",
      "ci_low": "0.8102564102564103",
      "ci_high": "0.8836633663366337",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.8259587020648967",
      "ci_low": "0.7840218423551757",
      "ci_high": "0.8647906357061286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1475",
      "ci_low": "0.11369509043927649",
      "ci_high": "0.18357607796597977",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.1575",
      "ci_low": "0.12339067507131035",
      "ci_high": "0.19444444444444445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "79.52499999999999",
      "ci_low": "76.25972157339064",
      "ci_high": "82.70623926116838",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "63.775000000000006",
      "ci_low": "57.517767540751244",
      "ci_high": "69.64965241721472",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.5473418155271824",
      "ci_low": "0.5095120118424594",
      "ci_high": "0.56817361341227",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.7525949963498759",
      "ci_low": "0.7005790162833817",
      "ci_high": "0.7812387184418713",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5794871794871795",
      "ci_low": "0.5274151436031331",
      "ci_high": "0.6308900523560209",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.6974358974358974",
      "ci_low": "0.6505102040816326",
      "ci_high": "0.7435233160621761",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.8308823529411765",
      "ci_low": "0.7829181494661922",
      "ci_high": "0.8745868644067795",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.11794871794871795",
      "ci_low": "0.08638743455497382",
      "ci_high": "0.15151700377861746",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.18717948717948718",
      "ci_low": "0.14786751361161524",
      "ci_high": "0.22894736842105262",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "72.64102564102564",
      "ci_low": "68.88607259174698",
      "ci_high": "76.29727880362562",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "53.92307692307693",
      "ci_low": "47.00232638571671",
      "ci_high": "60.52238805970149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5865229883529476",
      "ci_low": "0.5337536679544457",
      "ci_high": "0.6122990938092412",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.6466791922865834",
      "ci_low": "0.5884976338984914",
      "ci_high": "0.6750990008665992",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7142857142857143",
      "ci_low": "0.662379421221865",
      "ci_high": "0.7643312101910829",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8305647840531561",
      "ci_low": "0.7878787878787878",
      "ci_high": "0.8712895164114481",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.86",
      "ci_low": "0.816",
      "ci_high": "0.9012358925449054",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.11627906976744186",
      "ci_low": "0.08156028368794327",
      "ci_high": "0.1541404757875395",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.15946843853820597",
      "ci_low": "0.11960132890365449",
      "ci_high": "0.20202106683271065",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "79.5016611295681",
      "ci_low": "75.55185366958622",
      "ci_high": "83.32258822934573",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "63.55481727574751",
      "ci_low": "56.60837616447157",
      "ci_high": "70.43325223669753",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6769217582523076",
      "ci_low": "0.5839404666649041",
      "ci_high": "0.6957637173477168",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7276908901212307",
      "ci_low": "0.627736001664772",
      "ci_high": "0.7479459961487955",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.5333333333333333",
      "ci_low": "0.4673913043478261",
      "ci_high": "0.597070914696814",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6555555555555556",
      "ci_low": "0.5938566552901023",
      "ci_high": "0.7153572856064306",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.8135593220338984",
      "ci_low": "0.7513506363506364",
      "ci_high": "0.8704663212435233",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.12222222222222222",
      "ci_low": "0.08424908424908426",
      "ci_high": "0.16363636363636364",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.17777777777777778",
      "ci_low": "0.13026779677388195",
      "ci_high": "0.22756529850746265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "70.88888888888889",
      "ci_low": "66.26921641791046",
      "ci_high": "75.37641824249165",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "53.11111111111111",
      "ci_low": "44.68733377659574",
      "ci_high": "61.31098462448255",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.4690935915526805",
      "ci_low": "0.4201643423942415",
      "ci_high": "0.49685700795268667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6080842853460673",
      "ci_low": "0.5446574808814242",
      "ci_high": "0.6440738991979272",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.6712328767123288",
      "ci_low": "0.6051464592274678",
      "ci_high": "0.7339449541284404",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.8401826484018264",
      "ci_low": "0.7892376681614349",
      "ci_high": "0.8873269922565697",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7989130434782609",
      "ci_low": "0.7362637362637363",
      "ci_high": "0.8571428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.1689497716894977",
      "ci_low": "0.11946902654867257",
      "ci_high": "0.2231769236664623",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.182648401826484",
      "ci_low": "0.13157585866073634",
      "ci_high": "0.2391304347826087",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "77.94520547945206",
      "ci_low": "73.21577929714752",
      "ci_high": "82.46234797174276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "59.68036529680365",
      "ci_low": "50.701652386780914",
      "ci_high": "68.16922222993131",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.5658356431450248",
      "ci_low": "0.47481345573261163",
      "ci_high": "0.5879302948300407",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.7373009895526079",
      "ci_low": "0.6186963211061303",
      "ci_high": "0.7660909902330834",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7123287671232876",
      "ci_low": "0.6598639455782312",
      "ci_high": "0.7633333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8287671232876712",
      "ci_low": "0.7849829351535836",
      "ci_high": "0.8702545146239253",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.859504132231405",
      "ci_low": "0.8143459915611815",
      "ci_high": "0.9015748031496063",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.11643835616438356",
      "ci_low": "0.081078701941378",
      "ci_high": "0.1548833404467207",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.16095890410958905",
      "ci_low": "0.12027437035103745",
      "ci_high": "0.2048611111111111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "79.31506849315069",
      "ci_low": "75.33101045296166",
      "ci_high": "83.20286562107101",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "63.219178082191775",
      "ci_low": "56.10261317861393",
      "ci_high": "70.22879579463364",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6810733733519438",
      "ci_low": "0.5771786987928431",
      "ci_high": "0.6964444317720854",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.7309080104264762",
      "ci_low": "0.6194112865093926",
      "ci_high": "0.7474037804383354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.6894736842105263",
      "ci_low": "0.616580310880829",
      "ci_high": "0.7580645161290323",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8157894736842105",
      "ci_low": "0.7527472527472527",
      "ci_high": "0.8735632183908046",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8451612903225807",
      "ci_low": "0.7826013013901213",
      "ci_high": "0.9019607843137255",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.12631578947368421",
      "ci_low": "0.07936507936507936",
      "ci_high": "0.17857142857142858",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.15263157894736842",
      "ci_low": "0.10204081632653061",
      "ci_high": "0.20707070707070707",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "78.47368421052632",
      "ci_low": "72.82924365607292",
      "ci_high": "83.78537493579866",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "63.21052631578947",
      "ci_low": "53.6084315318976",
      "ci_high": "72.12465261422516",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.574218047518357",
      "ci_low": "0.4435896239218298",
      "ci_high": "0.5722025758938418",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.7002659116077526",
      "ci_low": "0.5409629560022314",
      "ci_high": "0.697808019382734",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.449438202247191",
      "ci_low": "0.3333333333333333",
      "ci_high": "0.5625151699029124",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6966292134831461",
      "ci_low": "0.5858585858585859",
      "ci_high": "0.7959183673469388",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6451612903225806",
      "ci_low": "0.5161290322580645",
      "ci_high": "0.7714285714285715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.24719101123595505",
      "ci_low": "0.15492957746478872",
      "ci_high": "0.34285714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2696629213483146",
      "ci_low": "0.1724137931034483",
      "ci_high": "0.37209302325581395",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "64.71910112359551",
      "ci_low": "56.665929203539825",
      "ci_high": "72.31582156611039",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "37.752808988764045",
      "ci_low": "21.950191833379012",
      "ci_high": "53.21051198257079",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.36465661531451005",
      "ci_low": "0.2022466856060606",
      "ci_high": "0.36738010687229433",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.530409622275651",
      "ci_low": "0.2941769972451791",
      "ci_high": "0.5343710645415192",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2222222222222222",
      "ci_high": "0.7777777777777778",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.4",
      "ci_high": "0.9166666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.75",
      "ci_low": "0.4",
      "ci_high": "1.0",
      "valid_resamples": "9998",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.16666666666666666",
      "ci_low": "0.0",
      "ci_high": "0.42857142857142855",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4166666666666667",
      "ci_low": "0.14285714285714285",
      "ci_high": "0.7",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "65.0",
      "ci_low": "45.33157894736842",
      "ci_high": "85.45454545454545",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "23.333333333333336",
      "ci_low": "-22.22916666666666",
      "ci_high": "70.66666666666666",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.40606060606060607",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.43636363636363634",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.4962962962962963",
      "ci_low": "0.1111111111111111",
      "ci_high": "0.5333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5771144278606966",
      "ci_low": "0.5053749142072752",
      "ci_high": "0.6481504745951981",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6865671641791045",
      "ci_low": "0.6197183098591549",
      "ci_high": "0.7512953367875648",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.8405797101449275",
      "ci_low": "0.7755065069506063",
      "ci_high": "0.8986507868628122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.10945273631840796",
      "ci_low": "0.06862745098039216",
      "ci_high": "0.15463917525773196",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.14925373134328357",
      "ci_low": "0.1",
      "ci_high": "0.20297029702970298",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "74.72636815920399",
      "ci_low": "69.81106261147094",
      "ci_high": "79.34589953271028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "59.80099502487562",
      "ci_low": "50.738825702694875",
      "ci_high": "68.2796118951613",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.5400642466643029",
      "ci_low": "0.4521432730938782",
      "ci_high": "0.5663076518607056",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6600785237008148",
      "ci_low": "0.5526195560036289",
      "ci_high": "0.6921537967186402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "0.0",
      "ci_high": "0.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "100.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "10.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "-90.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.5285714285714286",
      "ci_low": "0.3972602739726027",
      "ci_high": "0.6533333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.7857142857142857",
      "ci_low": "0.6774193548387096",
      "ci_high": "0.8783783783783784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6727272727272727",
      "ci_low": "0.5357142857142857",
      "ci_high": "0.8030365126676601",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2571428571428571",
      "ci_low": "0.15492957746478872",
      "ci_high": "0.3670886075949367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2571428571428571",
      "ci_low": "0.15492957746478872",
      "ci_high": "0.3670886075949367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "69.28571428571429",
      "ci_low": "60.0",
      "ci_high": "77.91680194805195",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "43.57142857142857",
      "ci_low": "26.19701542588867",
      "ci_high": "60.256493506493506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.33476406757034793",
      "ci_low": "0.2567505630334578",
      "ci_high": "0.38045735685578574",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6695281351406959",
      "ci_low": "0.5135011260669156",
      "ci_high": "0.7609147137115715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7111111111111111",
      "ci_low": "0.5714285714285714",
      "ci_high": "0.8387096774193549",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8444444444444444",
      "ci_low": "0.7316838649155722",
      "ci_high": "0.9423076923076923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8421052631578947",
      "ci_low": "0.7096774193548387",
      "ci_high": "0.95",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.13333333333333333",
      "ci_low": "0.041666666666666664",
      "ci_high": "0.25",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.17777777777777778",
      "ci_low": "0.07315964523281597",
      "ci_high": "0.30303030303030304",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "80.22222222222223",
      "ci_low": "69.44444444444444",
      "ci_high": "89.55579710144927",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "62.44444444444444",
      "ci_low": "42.14161129568107",
      "ci_high": "80.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.6532738095238095",
      "ci_low": "0.3693271982706126",
      "ci_high": "0.6641930465367964",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.8040293040293041",
      "ci_low": "0.454556551717677",
      "ci_high": "0.8174683649683648",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.5288888888888889",
      "ci_low": "0.4553182840339042",
      "ci_high": "0.6009179695912478",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6577777777777778",
      "ci_low": "0.5896200667380737",
      "ci_high": "0.7239819004524887",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.8040540540540541",
      "ci_low": "0.7346938775510204",
      "ci_high": "0.8675496688741722",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.1288888888888889",
      "ci_low": "0.08583690987124463",
      "ci_high": "0.17573589705851425",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.18222222222222223",
      "ci_low": "0.12931034482758622",
      "ci_high": "0.2381003039513677",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "70.57777777777778",
      "ci_low": "65.38109055281305",
      "ci_high": "75.55555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "52.35555555555556",
      "ci_low": "42.90310892940626",
      "ci_high": "61.320911204506835",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.47379955049127975",
      "ci_low": "0.4114907941298737",
      "ci_high": "0.5038908344328417",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.614184602488696",
      "ci_low": "0.533413992390577",
      "ci_high": "0.6531918224129427",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5774647887323944",
      "ci_low": "0.45761501210653754",
      "ci_high": "0.6935483870967742",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6619718309859155",
      "ci_low": "0.5492888704777686",
      "ci_high": "0.7727272727272727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8723404255319149",
      "ci_low": "0.7659574468085106",
      "ci_high": "0.96",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.08450704225352113",
      "ci_low": "0.02631578947368421",
      "ci_high": "0.15584415584415584",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.1267605633802817",
      "ci_low": "0.05405405405405406",
      "ci_high": "0.2112676056338028",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "72.95774647887325",
      "ci_low": "64.399375",
      "ci_high": "81.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "60.28169014084507",
      "ci_low": "45.59888888888889",
      "ci_high": "73.6231884057971",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6063775510204081",
      "ci_low": "0.3879818594104309",
      "ci_high": "0.6083189677386106",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6288359788359789",
      "ci_low": "0.40235155790711347",
      "ci_high": "0.6308492998770777",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7142857142857143",
      "ci_low": "0.662379421221865",
      "ci_high": "0.7643312101910829",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8305647840531561",
      "ci_low": "0.7878787878787878",
      "ci_high": "0.8712895164114481",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.86",
      "ci_low": "0.816",
      "ci_high": "0.9012358925449054",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.11627906976744186",
      "ci_low": "0.08156028368794327",
      "ci_high": "0.1541404757875395",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.15946843853820597",
      "ci_low": "0.11960132890365449",
      "ci_high": "0.20202106683271065",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "79.5016611295681",
      "ci_low": "75.55185366958622",
      "ci_high": "83.32258822934573",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "63.55481727574751",
      "ci_low": "56.60837616447157",
      "ci_high": "70.43325223669753",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6769217582523076",
      "ci_low": "0.5839404666649041",
      "ci_high": "0.6957637173477168",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.7276908901212307",
      "ci_low": "0.627736001664772",
      "ci_high": "0.7479459961487955",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.7948717948717948",
      "ci_low": "0.6941176470588235",
      "ci_high": "0.8852459016393442",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.9358974358974359",
      "ci_low": "0.8765432098765432",
      "ci_high": "0.9863013698630136",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8493150684931506",
      "ci_low": "0.75",
      "ci_high": "0.9342105263157895",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.14102564102564102",
      "ci_low": "0.0625",
      "ci_high": "0.234375",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.15384615384615385",
      "ci_low": "0.07228915662650602",
      "ci_high": "0.25",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "85.76923076923077",
      "ci_low": "78.43727678571429",
      "ci_high": "92.27283978397838",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "70.38461538461539",
      "ci_low": "54.53280487804878",
      "ci_high": "84.21052631578947",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5269910873440284",
      "ci_low": "0.40445851370851366",
      "ci_high": "0.5544705069124424",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8234235739750445",
      "ci_low": "0.6319664276695526",
      "ci_high": "0.8663601670506913",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7376543209876543",
      "ci_low": "0.6881347757370975",
      "ci_high": "0.7850746268656716",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.8611111111111112",
      "ci_low": "0.8207540563395169",
      "ci_high": "0.8993906949054697",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.8566308243727598",
      "ci_low": "0.8136185529433206",
      "ci_high": "0.896797153024911",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.12345679012345678",
      "ci_low": "0.0884124701854274",
      "ci_high": "0.16088418724391745",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.13271604938271606",
      "ci_low": "0.09687477742165243",
      "ci_high": "0.17251485342749845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "82.09876543209876",
      "ci_low": "78.52557692307693",
      "ci_high": "85.49394371713102",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "68.82716049382717",
      "ci_low": "62.12941919191919",
      "ci_high": "74.98444435254872",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5835279762023525",
      "ci_low": "0.5304215943957684",
      "ci_high": "0.6049760598081002",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.7831559680610518",
      "ci_low": "0.7118816135311629",
      "ci_high": "0.8119415539529766",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.5394736842105263",
      "ci_low": "0.41975308641975306",
      "ci_high": "0.654320987654321",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.7894736842105263",
      "ci_low": "0.6933333333333334",
      "ci_high": "0.8809523809523809",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6833333333333333",
      "ci_low": "0.5535714285714286",
      "ci_high": "0.8028169014084507",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.15384615384615385",
      "ci_high": "0.35714285714285715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2631578947368421",
      "ci_low": "0.1684925093632959",
      "ci_high": "0.36709506587445095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "68.55263157894737",
      "ci_low": "59.66657303370787",
      "ci_high": "76.90171274560946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "42.23684210526316",
      "ci_low": "24.444444444444446",
      "ci_high": "58.3566352739726",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.3525867136978248",
      "ci_low": "0.23130200817700816",
      "ci_high": "0.3825755571588904",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.47599206349206347",
      "ci_low": "0.312257711038961",
      "ci_high": "0.5164770021645022",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.7142857142857143",
      "ci_low": "0.6224489795918368",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8285714285714286",
      "ci_low": "0.7563025210084033",
      "ci_high": "0.897196261682243",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8620689655172413",
      "ci_low": "0.7821782178217822",
      "ci_high": "0.9325842696629213",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.11428571428571428",
      "ci_low": "0.05504587155963303",
      "ci_high": "0.1810344827586207",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.1523809523809524",
      "ci_low": "0.08256880733944955",
      "ci_high": "0.2298879310344827",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "81.14285714285714",
      "ci_low": "74.73672462406016",
      "ci_high": "87.14285714285714",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "65.9047619047619",
      "ci_low": "52.81811562224184",
      "ci_high": "77.98124999999999",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.661547217954781",
      "ci_low": "0.48494878400077984",
      "ci_high": "0.6728746328076685",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.7718050876139112",
      "ci_low": "0.565773581334243",
      "ci_high": "0.7850204049422798",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5298245614035088",
      "ci_low": "0.46964398946418556",
      "ci_high": "0.5907476245845225",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.6491228070175439",
      "ci_low": "0.5912162162162162",
      "ci_high": "0.706081383803158",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.8162162162162162",
      "ci_low": "0.7566103297968583",
      "ci_high": "0.8715083798882681",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.11929824561403508",
      "ci_low": "0.08243727598566308",
      "ci_high": "0.15916955017301038",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.2",
      "ci_low": "0.15441176470588236",
      "ci_high": "0.24919019933554787",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "69.50877192982456",
      "ci_low": "64.98305084745763",
      "ci_high": "73.9423076923077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "49.50877192982456",
      "ci_low": "41.517933868858584",
      "ci_high": "57.152126593649584",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5333835274757305",
      "ci_low": "0.4552077201535471",
      "ci_high": "0.5603667026108957",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5607365288847423",
      "ci_low": "0.47855170580244705",
      "ci_high": "0.5891034565909417",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8653061224489796",
      "ci_low": "0.8189646303391515",
      "ci_high": "0.9082578822963824",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9510204081632653",
      "ci_low": "0.9182128092552236",
      "ci_high": "0.9795081967213115",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9098712446351931",
      "ci_low": "0.8709677419354839",
      "ci_high": "0.9444444444444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.08571428571428572",
      "ci_low": "0.05263157894736842",
      "ci_high": "0.12244897959183673",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.08979591836734693",
      "ci_low": "0.05622489959839357",
      "ci_high": "0.12766191770550187",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "92.65306122448979",
      "ci_low": "90.08471916885327",
      "ci_high": "95.04467357470512",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "83.6734693877551",
      "ci_low": "77.59821110632157",
      "ci_high": "89.13056897477186",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.7473849382940291",
      "ci_low": "0.622920330821473",
      "ci_high": "0.7324318183425803",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.8360577275831513",
      "ci_low": "0.6968261327833427",
      "ci_high": "0.8193305086544119",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7153846153846154",
      "ci_low": "0.6363636363636364",
      "ci_high": "0.7950868077965664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.823076923076923",
      "ci_low": "0.7534223744292238",
      "ci_high": "0.8888888888888888",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.8691588785046729",
      "ci_low": "0.8021978021978022",
      "ci_high": "0.9304347826086956",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.1076923076923077",
      "ci_low": "0.05737704918032787",
      "ci_high": "0.16393442622950818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.15384615384615385",
      "ci_low": "0.09401709401709402",
      "ci_high": "0.21804511278195488",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "84.3076923076923",
      "ci_low": "79.53114442567568",
      "ci_high": "88.93959330143541",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "68.92307692307693",
      "ci_low": "59.08245494176589",
      "ci_high": "78.35708170254402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6379794973544973",
      "ci_low": "0.4953498151154401",
      "ci_high": "0.638349511721222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.6959776334776334",
      "ci_low": "0.5403816164895711",
      "ci_high": "0.6963812855140604",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.35526315789473684",
      "ci_low": "0.25",
      "ci_high": "0.4642857142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.5789473684210527",
      "ci_low": "0.46875",
      "ci_high": "0.6896673142302087",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.6136363636363636",
      "ci_low": "0.4666666666666667",
      "ci_high": "0.7560975609756098",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.2236842105263158",
      "ci_low": "0.13414634146341464",
      "ci_high": "0.3230769230769231",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.27631578947368424",
      "ci_low": "0.17857142857142858",
      "ci_high": "0.38095238095238093",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "50.26315789473684",
      "ci_low": "41.14285714285714",
      "ci_high": "59.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "22.63157894736842",
      "ci_low": "6.41970702045329",
      "ci_high": "38.51862373737374",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.3902964959568733",
      "ci_low": "0.18658243486073675",
      "ci_high": "0.34380091344713987",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.4496894409937888",
      "ci_low": "0.21497541407867496",
      "ci_high": "0.3961184437543133",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2755102040816326",
      "ci_low": "0.18811669628501312",
      "ci_high": "0.3695652173913043",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4897959183673469",
      "ci_low": "0.38823529411764707",
      "ci_high": "0.5904761904761905",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.5625",
      "ci_low": "0.4166666666666667",
      "ci_high": "0.7045454545454546",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.21428571428571427",
      "ci_low": "0.13592233009708737",
      "ci_high": "0.29896907216494845",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.29591836734693877",
      "ci_low": "0.20731432652164358",
      "ci_high": "0.38636363636363635",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "45.102040816326536",
      "ci_low": "37.04081632653061",
      "ci_high": "53.54183467741936",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "15.510204081632653",
      "ci_low": "1.5211956521739136",
      "ci_high": "29.487499999999997",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.23257142857142857",
      "ci_low": "0.13551338383838382",
      "ci_high": "0.2434307692307692",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.25279503105590057",
      "ci_low": "0.14729715634606938",
      "ci_high": "0.26459866220735784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6915887850467289",
      "ci_low": "0.603448275862069",
      "ci_high": "0.7764705882352941",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.822429906542056",
      "ci_low": "0.7474720893141946",
      "ci_high": "0.8910891089108911",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.8409090909090909",
      "ci_low": "0.759493670886076",
      "ci_high": "0.9139784946236559",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.1308411214953271",
      "ci_low": "0.06930693069306931",
      "ci_high": "0.19820013979496737",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.14953271028037382",
      "ci_low": "0.0847457627118644",
      "ci_high": "0.22",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "79.4392523364486",
      "ci_low": "73.13029891304348",
      "ci_high": "85.37831286360698",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "64.48598130841121",
      "ci_low": "52.345957613814754",
      "ci_high": "75.71428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.6417898193760262",
      "ci_low": "0.4235570607553366",
      "ci_high": "0.590068076081007",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.7298786181139122",
      "ci_low": "0.48169234360410823",
      "ci_high": "0.6710578120136943",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5413533834586466",
      "ci_low": "0.45384615384615384",
      "ci_high": "0.6283249051833122",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6766917293233082",
      "ci_low": "0.5948261775974758",
      "ci_high": "0.7575757575757576",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.8",
      "ci_low": "0.7093023255813954",
      "ci_high": "0.8829787234042553",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.13533834586466165",
      "ci_low": "0.0782608695652174",
      "ci_high": "0.2",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.21052631578947367",
      "ci_low": "0.1399965034965035",
      "ci_high": "0.2868429436126152",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "72.4812030075188",
      "ci_low": "66.52882304671473",
      "ci_high": "78.05975934282078",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "51.42857142857143",
      "ci_low": "39.59401544401545",
      "ci_high": "62.45904892011449",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5442528060175119",
      "ci_low": "0.37526196842373316",
      "ci_high": "0.5285456547572916",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5664672062631246",
      "ci_low": "0.3905787834614366",
      "ci_high": "0.5501189467882014",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Gemini 3.8 HIGH media / HIGH thinking",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.48951048951048953",
      "ci_low": "0.4397313015972721",
      "ci_high": "0.5395348837209303",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8717948717948718",
      "ci_low": "0.8400954653937948",
      "ci_high": "0.9026144159139113",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.5614973262032086",
      "ci_low": "0.5079787234042553",
      "ci_high": "0.6143617021276596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.3822843822843823",
      "ci_low": "0.3348309624056404",
      "ci_high": "0.43137254901960786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4125874125874126",
      "ci_low": "0.3642852999668545",
      "ci_high": "0.46153846153846156",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "63.4965034965035",
      "ci_low": "59.58311170212765",
      "ci_high": "67.39606757151932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "22.237762237762237",
      "ci_low": "14.608412744102898",
      "ci_high": "29.721195567348207",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.27580270810430607",
      "ci_low": "0.23754784436190526",
      "ci_high": "0.2853211336224742",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.46708523146696995",
      "ci_low": "0.40229876867742026",
      "ci_high": "0.4832051456509644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.30193905817174516",
      "ci_low": "0.2532964485624485",
      "ci_high": "0.35278514588859416",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.8365650969529086",
      "ci_low": "0.7946416505791505",
      "ci_high": "0.8761329305135952",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3609271523178808",
      "ci_low": "0.30491803278688523",
      "ci_high": "0.41883116883116883",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5346260387811634",
      "ci_low": "0.4807692307692308",
      "ci_high": "0.5894751341162824",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.554016620498615",
      "ci_low": "0.5",
      "ci_high": "0.6073328432769768",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "47.146814404432135",
      "ci_low": "42.928889007928",
      "ci_high": "51.44785293645883",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "-8.254847645429363",
      "ci_low": "-15.965497080435144",
      "ci_high": "0.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.1464895552300676",
      "ci_low": "0.11064952326137786",
      "ci_high": "0.15225922164258712",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3078763533648877",
      "ci_low": "0.2325515404137433",
      "ci_high": "0.3200024319267933",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5866666666666667",
      "ci_low": "0.5345744680851063",
      "ci_high": "0.6381782945736434",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.9786666666666667",
      "ci_low": "0.9629629629629629",
      "ci_high": "0.9920429666119321",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5994550408719346",
      "ci_low": "0.5474254742547425",
      "ci_high": "0.6511643748811863",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.392",
      "ci_low": "0.3410841510955774",
      "ci_high": "0.4432724478415798",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.42133333333333334",
      "ci_low": "0.36923076923076925",
      "ci_high": "0.47342071859243323",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "73.81333333333333",
      "ci_low": "70.16215123479274",
      "ci_high": "77.35220399213671",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "31.68",
      "ci_low": "23.386603658536586",
      "ci_high": "39.94640191841095",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.33875402284119394",
      "ci_low": "0.2927426374648299",
      "ci_high": "0.3443524435912997",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5559040374829848",
      "ci_low": "0.4803981743012593",
      "ci_high": "0.5650911894831584",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1206896551724138",
      "ci_low": "0.07262569832402235",
      "ci_high": "0.17318435754189945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5402298850574713",
      "ci_low": "0.4634146341463415",
      "ci_high": "0.6170243161094224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.22340425531914893",
      "ci_low": "0.13829787234042554",
      "ci_high": "0.3113207547169811",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.41954022988505746",
      "ci_low": "0.3465909090909091",
      "ci_high": "0.49404761904761907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.4367816091954023",
      "ci_low": "0.3619631901840491",
      "ci_high": "0.5126590310351474",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "23.218390804597703",
      "ci_low": "18.125",
      "ci_high": "28.54670425889605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "-20.45977011494253",
      "ci_low": "-28.80250578388677",
      "ci_high": "-11.904511278195493",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.07115812536373284",
      "ci_low": "0.03657070538495772",
      "ci_high": "0.08156063303259564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.09888207031064172",
      "ci_low": "0.05081903215831787",
      "ci_high": "0.11333750304529523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.32083333333333336",
      "ci_low": "0.2616033755274262",
      "ci_high": "0.38257658102766795",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.8916666666666667",
      "ci_low": "0.8508754244482173",
      "ci_high": "0.9297520661157025",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3598130841121495",
      "ci_low": "0.2949293985584308",
      "ci_high": "0.42660550458715596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5708333333333333",
      "ci_low": "0.5061728395061729",
      "ci_high": "0.6347860593512767",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5958333333333333",
      "ci_low": "0.5321078320603787",
      "ci_high": "0.6600790513833992",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "51.833333333333336",
      "ci_low": "46.94684119960669",
      "ci_high": "56.78862720966659",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "-7.75",
      "ci_low": "-17.3556479086493",
      "ci_high": "2.2326513590422907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2151797535797536",
      "ci_low": "0.15615385336885335",
      "ci_high": "0.21379193192772142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3404742936388506",
      "ci_low": "0.24707888191274266",
      "ci_high": "0.33827837330335664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.455",
      "ci_low": "0.40540540540540543",
      "ci_high": "0.50495111078272",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.875",
      "ci_low": "0.8414322250639387",
      "ci_high": "0.9066994942608523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.52",
      "ci_low": "0.4668587896253602",
      "ci_high": "0.5738185781760929",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.42",
      "ci_low": "0.36939313984168864",
      "ci_high": "0.469334693877551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4525",
      "ci_low": "0.40245962672855623",
      "ci_high": "0.5024878731343284",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "60.099999999999994",
      "ci_low": "55.97328909952606",
      "ci_high": "64.11488330696201",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "14.850000000000001",
      "ci_low": "6.828572780154917",
      "ci_high": "22.700588957941896",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2042805257464443",
      "ci_low": "0.1858802338328159",
      "ci_high": "0.21486251008167706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4953802749351276",
      "ci_low": "0.4507595670445786",
      "ci_high": "0.5210415869480669",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.35128205128205126",
      "ci_low": "0.3010471204188482",
      "ci_high": "0.4036001432626392",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.8358974358974359",
      "ci_low": "0.7964824120603015",
      "ci_high": "0.8737373737373737",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.42024539877300615",
      "ci_low": "0.3630947802197802",
      "ci_high": "0.47896440129449835",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.4846153846153846",
      "ci_low": "0.43116883116883115",
      "ci_high": "0.5382716049382716",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5025641025641026",
      "ci_low": "0.4487798121748531",
      "ci_high": "0.5566502463054187",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "51.84615384615385",
      "ci_low": "47.63769008835728",
      "ci_high": "56.153846153846146",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "1.5897435897435896",
      "ci_low": "-6.310468785428161",
      "ci_high": "9.639061080835601",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.1879247540821163",
      "ci_low": "0.1609243544252662",
      "ci_high": "0.2028975237292849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.3806680916022355",
      "ci_low": "0.3259749743486162",
      "ci_high": "0.41099754806701305",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5780730897009967",
      "ci_low": "0.5218818620992534",
      "ci_high": "0.6330291729918817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8471760797342193",
      "ci_low": "0.8049623767514271",
      "ci_high": "0.887720193340494",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6823529411764706",
      "ci_low": "0.6245349599054253",
      "ci_high": "0.7381822083495903",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2691029900332226",
      "ci_low": "0.21981339094342192",
      "ci_high": "0.32094594594594594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.32558139534883723",
      "ci_low": "0.2743055555555556",
      "ci_high": "0.37888362209872506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "65.98006644518273",
      "ci_low": "61.170740137642866",
      "ci_high": "70.71428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "33.421926910299",
      "ci_low": "24.934640522875817",
      "ci_high": "41.61491216836354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.3747442942594927",
      "ci_low": "0.31435163568859653",
      "ci_high": "0.3865324167486628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6183280855281629",
      "ci_low": "0.5186801988861843",
      "ci_high": "0.6377784876352937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2037037037037037",
      "ci_low": "0.1513918495758663",
      "ci_high": "0.26071428571428573",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.837037037037037",
      "ci_low": "0.7872340425531915",
      "ci_high": "0.8825910931174089",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.24336283185840707",
      "ci_low": "0.18260779361356666",
      "ci_high": "0.3094170403587444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6333333333333333",
      "ci_low": "0.5682656826568265",
      "ci_high": "0.694980694980695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6333333333333333",
      "ci_low": "0.5682656826568265",
      "ci_high": "0.694980694980695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "41.0",
      "ci_low": "36.21205283885789",
      "ci_high": "45.89745762711864",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "-22.333333333333336",
      "ci_low": "-31.31317500262137",
      "ci_high": "-12.941067133172973",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.07913757879966443",
      "ci_low": "0.06094051623524625",
      "ci_high": "0.09021121552442732",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.21982660777684565",
      "ci_low": "0.16927921176457286",
      "ci_high": "0.2505867097900759",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.410958904109589",
      "ci_low": "0.3417024145115823",
      "ci_high": "0.4772727272727273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.8904109589041096",
      "ci_low": "0.8471615720524017",
      "ci_high": "0.930739421868454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.46153846153846156",
      "ci_low": "0.38678758996304224",
      "ci_high": "0.532608695652174",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4794520547945205",
      "ci_low": "0.4131455399061033",
      "ci_high": "0.5485237188152882",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4931506849315068",
      "ci_low": "0.4266666666666667",
      "ci_high": "0.5627906976744186",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "60.86757990867579",
      "ci_low": "55.575126422250314",
      "ci_high": "66.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "11.552511415525116",
      "ci_low": "1.0253623188405798",
      "ci_high": "21.75008886255924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.26716329707618997",
      "ci_low": "0.20561165159729788",
      "ci_high": "0.2835337488332451",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.47765559174227906",
      "ci_low": "0.3676087104315326",
      "ci_high": "0.5069239751867107",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.5821917808219178",
      "ci_low": "0.5253595740008784",
      "ci_high": "0.6382270713818894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8424657534246576",
      "ci_low": "0.8",
      "ci_high": "0.8841085744161729",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6910569105691057",
      "ci_low": "0.6322314049586777",
      "ci_high": "0.7473047985687357",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.2602739726027397",
      "ci_low": "0.21111014994232988",
      "ci_high": "0.312280701754386",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.3184931506849315",
      "ci_low": "0.2664359861591695",
      "ci_high": "0.3727598566308244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "66.26712328767124",
      "ci_low": "61.3604710095959",
      "ci_high": "71.07269338844894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "34.417808219178085",
      "ci_low": "25.910468779535",
      "ci_high": "42.64826084344586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.38590240723266706",
      "ci_low": "0.3174225095315988",
      "ci_high": "0.39435467263546026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6212087531062446",
      "ci_low": "0.5109728202215981",
      "ci_high": "0.6348148388765946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4052631578947368",
      "ci_low": "0.3315500755638224",
      "ci_high": "0.4759672619047618",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8736842105263158",
      "ci_low": "0.8163237727523442",
      "ci_high": "0.9257149929278642",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.463855421686747",
      "ci_low": "0.38460264900662255",
      "ci_high": "0.5403726708074534",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.46842105263157896",
      "ci_low": "0.39654917363803305",
      "ci_high": "0.5454545454545454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4842105263157895",
      "ci_low": "0.4111111111111111",
      "ci_high": "0.5612244897959183",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "59.21052631578948",
      "ci_low": "53.238595310036786",
      "ci_high": "64.79606991311377",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "10.789473684210527",
      "ci_low": "-0.621126785275643",
      "ci_high": "21.571482958665698",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.23327391256692753",
      "ci_low": "0.16734852808911027",
      "ci_high": "0.24091852449307474",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4153413565216027",
      "ci_low": "0.29796201342695244",
      "ci_high": "0.42895249482913306",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.24719101123595505",
      "ci_low": "0.14633461222252672",
      "ci_high": "0.3561683199496142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.8651685393258427",
      "ci_low": "0.7926829268292683",
      "ci_high": "0.9322091158955564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.2857142857142857",
      "ci_low": "0.171875",
      "ci_high": "0.40625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6179775280898876",
      "ci_low": "0.5063291139240507",
      "ci_high": "0.7241467276565797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6179775280898876",
      "ci_low": "0.5063291139240507",
      "ci_high": "0.7241467276565797",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "47.865168539325836",
      "ci_low": "39.404582616179",
      "ci_high": "56.78607142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "-13.932584269662922",
      "ci_low": "-30.0",
      "ci_high": "3.1959793814432973",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.07926587301587301",
      "ci_low": "0.043892045454545454",
      "ci_high": "0.09858506944444444",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.17294372294372293",
      "ci_low": "0.09576446280991736",
      "ci_high": "0.21509469696969696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.75",
      "ci_low": "0.5",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.6666666666666666",
      "valid_resamples": "9999",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2222222222222222",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.2222222222222222",
      "ci_high": "0.8",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "40.0",
      "ci_low": "17.5",
      "ci_high": "67.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "-10.0",
      "ci_low": "-48.891666666666666",
      "ci_high": "36.666666666666664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.13333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.13333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.2222222222222222",
      "ci_low": "0.0",
      "ci_high": "0.2222222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.21890547263681592",
      "ci_low": "0.16097560975609757",
      "ci_high": "0.2803738317757009",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.8557213930348259",
      "ci_low": "0.8040201005025126",
      "ci_high": "0.9030612244897959",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.2558139534883721",
      "ci_low": "0.18974137931034482",
      "ci_high": "0.3258426966292135",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6368159203980099",
      "ci_low": "0.5666625615763546",
      "ci_high": "0.7035175879396985",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6368159203980099",
      "ci_low": "0.5666625615763546",
      "ci_high": "0.7035175879396985",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "42.636815920398014",
      "ci_low": "37.393435686653774",
      "ci_high": "47.857522796352576",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "-21.044776119402986",
      "ci_low": "-30.9712133407113",
      "ci_high": "-10.811752649576286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.10154783075906071",
      "ci_low": "0.0700121823670379",
      "ci_high": "0.11280922090559384",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.2482280307443706",
      "ci_low": "0.1711408902305371",
      "ci_high": "0.2757558733247849",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "10.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "-90.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "-80.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "-80.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2714285714285714",
      "ci_low": "0.15554597701149425",
      "ci_high": "0.3918918918918919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.8857142857142857",
      "ci_low": "0.8051948051948052",
      "ci_high": "0.9552238805970149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3064516129032258",
      "ci_low": "0.17855158730158732",
      "ci_high": "0.43859649122807015",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6142857142857143",
      "ci_low": "0.4918032786885246",
      "ci_high": "0.7391304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6142857142857143",
      "ci_low": "0.4918032786885246",
      "ci_high": "0.7391304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "50.42857142857143",
      "ci_low": "41.14251152073732",
      "ci_high": "59.63652597402598",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "-11.0",
      "ci_low": "-29.51229346970889",
      "ci_high": "8.197282730743312",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.09838243597834391",
      "ci_low": "0.06452431927975406",
      "ci_high": "0.12496990990916822",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3232565753574157",
      "ci_low": "0.21200847763347763",
      "ci_high": "0.41061541827298126",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.2682926829268293",
      "ci_high": "0.5277908805031445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7777777777777778",
      "ci_low": "0.65",
      "ci_high": "0.8947368421052632",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5142857142857142",
      "ci_low": "0.36363636363636365",
      "ci_high": "0.6585365853658537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.37777777777777777",
      "ci_low": "0.25",
      "ci_high": "0.5116279069767442",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.2727272727272727",
      "ci_high": "0.5365853658536586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "55.33333333333333",
      "ci_low": "43.61659574468086",
      "ci_high": "66.66836734693875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "15.333333333333334",
      "ci_low": "-3.043780193236715",
      "ci_high": "33.19205927051671",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.3279799565513851",
      "ci_low": "0.14681234335839602",
      "ci_high": "0.34365740740740736",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5298137759676222",
      "ci_low": "0.23715840080971665",
      "ci_high": "0.5551388888888887",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.20444444444444446",
      "ci_low": "0.14678899082568808",
      "ci_high": "0.2672822466578455",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.8355555555555556",
      "ci_low": "0.7800809128630706",
      "ci_high": "0.886077515979446",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.24468085106382978",
      "ci_low": "0.1774193548387097",
      "ci_high": "0.3177083333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6311111111111111",
      "ci_low": "0.5589519650655022",
      "ci_high": "0.6995515695067265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6311111111111111",
      "ci_low": "0.5589519650655022",
      "ci_high": "0.6995515695067265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "41.11111111111111",
      "ci_low": "35.8332345971564",
      "ci_high": "46.62222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "-22.0",
      "ci_low": "-32.008944468392144",
      "ci_high": "-11.585450421862173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.08174183284799899",
      "ci_low": "0.05899417235217674",
      "ci_high": "0.09370317946940768",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.21797822092799735",
      "ci_low": "0.15731779293913797",
      "ci_high": "0.24987514525175383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.28169014084507044",
      "ci_low": "0.1780821917808219",
      "ci_high": "0.390625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8873239436619719",
      "ci_low": "0.8088235294117647",
      "ci_high": "0.9558983375959077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.31746031746031744",
      "ci_low": "0.2033898305084746",
      "ci_high": "0.4383561643835616",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.49019607843137253",
      "ci_high": "0.71875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.49019607843137253",
      "ci_high": "0.71875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "46.90140845070423",
      "ci_low": "37.94498804087846",
      "ci_high": "55.91556481747629",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "-13.661971830985916",
      "ci_low": "-31.0356477166822",
      "ci_high": "4.69796650717702",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.20703463203463202",
      "ci_low": "0.11009920634920636",
      "ci_high": "0.22435749540863179",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.33738977072310405",
      "ci_low": "0.17942092886537334",
      "ci_high": "0.36561962214739996",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5780730897009967",
      "ci_low": "0.5218818620992534",
      "ci_high": "0.6330291729918817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8471760797342193",
      "ci_low": "0.8049623767514271",
      "ci_high": "0.887720193340494",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6823529411764706",
      "ci_low": "0.6245349599054253",
      "ci_high": "0.7381822083495903",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2691029900332226",
      "ci_low": "0.21981339094342192",
      "ci_high": "0.32094594594594594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.32558139534883723",
      "ci_low": "0.2743055555555556",
      "ci_high": "0.37888362209872506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "65.98006644518273",
      "ci_low": "61.170740137642866",
      "ci_high": "70.71428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "33.421926910299",
      "ci_low": "24.934640522875817",
      "ci_high": "41.61491216836354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.3747442942594927",
      "ci_low": "0.31435163568859653",
      "ci_high": "0.3865324167486628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6183280855281629",
      "ci_low": "0.5186801988861843",
      "ci_high": "0.6377784876352937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5384615384615384",
      "ci_low": "0.4177215189873418",
      "ci_high": "0.6565694255992763",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.9358974358974359",
      "ci_low": "0.8765432098765432",
      "ci_high": "0.9863013698630136",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5753424657534246",
      "ci_low": "0.45",
      "ci_high": "0.6935483870967742",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3974358974358974",
      "ci_low": "0.2857142857142857",
      "ci_high": "0.5172413793103449",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4230769230769231",
      "ci_low": "0.30985915492957744",
      "ci_high": "0.5454545454545454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "74.35897435897436",
      "ci_low": "66.58528866934239",
      "ci_high": "81.7074912891986",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "32.05128205128205",
      "ci_low": "14.507042253521128",
      "ci_high": "48.98876404494382",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.36635243841126197",
      "ci_low": "0.2716320831945832",
      "ci_high": "0.3977283133533133",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.549528657616893",
      "ci_low": "0.40744812479187476",
      "ci_high": "0.59659247002997",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5092592592592593",
      "ci_low": "0.45364238410596025",
      "ci_high": "0.5641891891891891",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.8734567901234568",
      "ci_low": "0.8368564371968766",
      "ci_high": "0.9093569459903535",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5830388692579506",
      "ci_low": "0.5243055555555556",
      "ci_high": "0.6411149825783972",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.36419753086419754",
      "ci_low": "0.3111111111111111",
      "ci_high": "0.4171789298338276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.404320987654321",
      "ci_low": "0.3501577287066246",
      "ci_high": "0.4591564188757925",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "64.62962962962963",
      "ci_low": "60.12617304525524",
      "ci_high": "68.94260466119982",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "24.19753086419753",
      "ci_low": "15.572243672677287",
      "ci_high": "32.477907854984885",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.2564240868637153",
      "ci_low": "0.2322738089112476",
      "ci_high": "0.2707035669158167",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5061001714415434",
      "ci_low": "0.4584351491669362",
      "ci_high": "0.5342833557549014",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2236842105263158",
      "ci_low": "0.1323529411764706",
      "ci_high": "0.32142857142857145",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.881578947368421",
      "ci_low": "0.8028169014084507",
      "ci_high": "0.95",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2537313432835821",
      "ci_low": "0.1506849315068493",
      "ci_high": "0.36363636363636365",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6578947368421053",
      "ci_low": "0.5443037974683544",
      "ci_high": "0.7666666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6578947368421053",
      "ci_low": "0.5443037974683544",
      "ci_high": "0.7666666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "40.78947368421052",
      "ci_low": "32.49932795698925",
      "ci_high": "49.48717948717949",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "-25.0",
      "ci_low": "-41.46693181818182",
      "ci_high": "-7.71345755693583",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.10329004329004329",
      "ci_low": "0.05166590909090908",
      "ci_high": "0.11227152406417111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2582251082251082",
      "ci_low": "0.12916477272727273",
      "ci_high": "0.28067881016042784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.42857142857142855",
      "ci_low": "0.3220290278589519",
      "ci_high": "0.5339805825242718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8666666666666667",
      "ci_low": "0.7982456140350878",
      "ci_high": "0.9285714285714286",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4945054945054945",
      "ci_low": "0.3764705882352941",
      "ci_high": "0.6104076479076477",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4380952380952381",
      "ci_low": "0.3333333333333333",
      "ci_high": "0.5447339246119731",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4380952380952381",
      "ci_low": "0.3333333333333333",
      "ci_high": "0.5447339246119731",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "51.96246444534742",
      "ci_high": "67.89486842105262",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "16.19047619047619",
      "ci_low": "0.295906481404444",
      "ci_high": "32.04352278545826",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.3007704642615128",
      "ci_low": "0.19920298270245912",
      "ci_high": "0.32366330650773373",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4260914910371432",
      "ci_low": "0.2822042254951504",
      "ci_high": "0.4585230175526228",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.32280701754385965",
      "ci_low": "0.2659923705136471",
      "ci_high": "0.3811188811188811",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.8245614035087719",
      "ci_low": "0.7755867523138554",
      "ci_high": "0.8709705518714173",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.39148936170212767",
      "ci_low": "0.3254897336293008",
      "ci_high": "0.4581714807436918",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5017543859649123",
      "ci_low": "0.44074074074074077",
      "ci_high": "0.5632224380767362",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5263157894736842",
      "ci_low": "0.4641509433962264",
      "ci_high": "0.5871886120996441",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "48.8421052631579",
      "ci_low": "43.96420454545455",
      "ci_high": "53.75922264045372",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "-3.7894736842105265",
      "ci_low": "-12.601073762838467",
      "ci_high": "5.3062562730010026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.1688338010912495",
      "ci_low": "0.13034962230182134",
      "ci_high": "0.18289251890430513",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.33333853035964645",
      "ci_low": "0.25735694659590364",
      "ci_high": "0.3610954860418332",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6857142857142857",
      "ci_low": "0.6234309623430963",
      "ci_high": "0.7449412874816523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9877551020408163",
      "ci_low": "0.97265625",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6942148760330579",
      "ci_low": "0.6322314049586777",
      "ci_high": "0.7543103448275862",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.3020408163265306",
      "ci_low": "0.24267782426778242",
      "ci_high": "0.3632657047193878",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.33877551020408164",
      "ci_low": "0.2767830674760023",
      "ci_high": "0.4016745342059149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "81.6734693877551",
      "ci_low": "77.64257782311664",
      "ci_high": "85.49025178253119",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "47.79591836734694",
      "ci_low": "37.89635796404502",
      "ci_high": "57.21168006545247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.43458188153310107",
      "ci_low": "0.36160535809021865",
      "ci_high": "0.4356614476300943",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6039951573849879",
      "ci_low": "0.5025701587016598",
      "ci_high": "0.6054955712825039",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.3153846153846154",
      "ci_high": "0.48717948717948717",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.9615384615384616",
      "ci_low": "0.9252336448598131",
      "ci_high": "0.9920634920634921",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.416",
      "ci_low": "0.3283582089552239",
      "ci_high": "0.5042735042735043",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5615384615384615",
      "ci_low": "0.4732824427480916",
      "ci_high": "0.648854961832061",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5769230769230769",
      "ci_low": "0.48836751510712323",
      "ci_high": "0.6623608534322817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "59.0",
      "ci_low": "52.55594274146905",
      "ci_high": "65.62021185884822",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "1.3076923076923077",
      "ci_low": "-12.636426767676769",
      "ci_high": "15.758463949843247",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.22110709262607994",
      "ci_low": "0.15301787084223792",
      "ci_high": "0.2298774362065501",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.3969877344877345",
      "ci_low": "0.2747366317394726",
      "ci_high": "0.4127344877344877",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.06578947368421052",
      "ci_low": "0.013513513513513514",
      "ci_high": "0.12698412698412698",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.4605263157894737",
      "ci_low": "0.34882840028188866",
      "ci_high": "0.5731747967479673",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.03125",
      "ci_high": "0.2702702702702703",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.39473684210526316",
      "ci_low": "0.287496408045977",
      "ci_high": "0.5061728395061729",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.42105263157894735",
      "ci_low": "0.3125",
      "ci_high": "0.5316455696202531",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "17.105263157894736",
      "ci_low": "10.833333333333332",
      "ci_high": "24.050632911392405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "-25.0",
      "ci_low": "-35.21790890269151",
      "ci_high": "-14.931169997754328",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.048205128205128206",
      "ci_low": "0.010256410256410256",
      "ci_high": "0.05846153846153846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.06811594202898551",
      "ci_low": "0.014492753623188404",
      "ci_high": "0.08260869565217391",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.16326530612244897",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.24691721132897596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.6020408163265306",
      "ci_low": "0.5",
      "ci_high": "0.7040816326530612",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2711864406779661",
      "ci_low": "0.15623263888888891",
      "ci_high": "0.39285714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4387755102040816",
      "ci_low": "0.33962264150943394",
      "ci_high": "0.5421686746987951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4489795918367347",
      "ci_low": "0.3474553969669938",
      "ci_high": "0.5505732630130702",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "27.95918367346939",
      "ci_low": "20.574630541871922",
      "ci_high": "35.94623346751006",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "-16.93877551020408",
      "ci_low": "-29.419552578361976",
      "ci_high": "-4.225273865414711",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.08222912353347137",
      "ci_low": "0.039195709684840124",
      "ci_high": "0.09879227053140097",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.12334368530020705",
      "ci_low": "0.058793564527260186",
      "ci_high": "0.14818840579710144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.3364485981308411",
      "ci_low": "0.24770642201834864",
      "ci_high": "0.42609369733737773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.897196261682243",
      "ci_low": "0.8365384615384616",
      "ci_high": "0.9516206575682381",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.375",
      "ci_low": "0.2792740414833438",
      "ci_high": "0.4717034131863472",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5607476635514018",
      "ci_low": "0.46609963797926607",
      "ci_high": "0.6551724137931034",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5794392523364486",
      "ci_low": "0.4891304347826087",
      "ci_high": "0.6728971962616822",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "54.48598130841121",
      "ci_low": "47.5",
      "ci_high": "61.61926910299003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "-3.4579439252336446",
      "ci_low": "-17.403935185185183",
      "ci_high": "10.488236018723818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2309206349206349",
      "ci_low": "0.14162012987012987",
      "ci_high": "0.22733333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.33958916900093367",
      "ci_low": "0.20826489686783803",
      "ci_high": "0.3343137254901961",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3082706766917293",
      "ci_low": "0.22880750605326877",
      "ci_high": "0.3944954128440367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.8872180451127819",
      "ci_low": "0.8309859154929577",
      "ci_high": "0.9396551724137931",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3474576271186441",
      "ci_low": "0.2601626016260163",
      "ci_high": "0.44118309485956536",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5789473684210527",
      "ci_low": "0.4895067698259188",
      "ci_high": "0.6642335766423357",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6090225563909775",
      "ci_low": "0.5220588235294118",
      "ci_high": "0.6936936936936937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "49.69924812030075",
      "ci_low": "43.09515977443609",
      "ci_high": "56.559999999999995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "-11.203007518796992",
      "ci_low": "-23.931623931623932",
      "ci_high": "2.307823129251699",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.19558103076621594",
      "ci_low": "0.12479618571938833",
      "ci_high": "0.19909069631528667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3233074182053774",
      "ci_low": "0.20629573557694805",
      "ci_high": "0.3291091102354738",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct first_pass",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.48951048951048953",
      "ci_low": "0.4397313015972721",
      "ci_high": "0.5395348837209303",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_answer_coverage",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.8741258741258742",
      "ci_low": "0.8426952033308618",
      "ci_high": "0.9047619047619048",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "species_attempted_accuracy",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.56",
      "ci_low": "0.5062034739454094",
      "ci_high": "0.6129055039183524",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "wrong_species_rate_all_events",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.38461538461538464",
      "ci_low": "0.33647058823529413",
      "ci_high": "0.43387470997679817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "any_rank_conflict_rate",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.4149184149184149",
      "ci_low": "0.36658157397767716",
      "ci_high": "0.46454902957786093",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "raw_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "63.4965034965035",
      "ci_low": "59.58311170212765",
      "ci_high": "67.39606757151932",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "adjusted_tis_pct",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "22.004662004662006",
      "ci_low": "14.315357505386144",
      "ci_high": "29.542439228193967",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.27320079576369943",
      "ci_low": "0.23530682696226463",
      "ci_high": "0.28262942481471504",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "day",
      "metric": "reference_species_macro_f1",
      "events": "429",
      "species": "62",
      "source_groups": "395",
      "method": "source_cluster",
      "estimate": "0.46708523146696995",
      "ci_low": "0.40229876867742026",
      "ci_high": "0.4832051456509644",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.3047091412742382",
      "ci_low": "0.2553191489361702",
      "ci_high": "0.355747044580544",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_answer_coverage",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.8448753462603878",
      "ci_low": "0.8033696193470073",
      "ci_high": "0.8834701820029564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "species_attempted_accuracy",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.36065573770491804",
      "ci_low": "0.30475875118259227",
      "ci_high": "0.41818478906714196",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5401662049861495",
      "ci_low": "0.48633879781420764",
      "ci_high": "0.5945223991114402",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.5595567867036011",
      "ci_low": "0.5051175548589342",
      "ci_high": "0.6135709975864843",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "raw_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "47.50692520775623",
      "ci_low": "43.33310055865921",
      "ci_high": "51.79561814396669",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "adjusted_tis_pct",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "-8.448753462603879",
      "ci_low": "-16.25662956379241",
      "ci_high": "-0.1408155637737041",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.147244129838632",
      "ci_low": "0.11118732279369133",
      "ci_high": "0.15303127974848377",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "time_of_day",
      "stratum": "night_ir",
      "metric": "reference_species_macro_f1",
      "events": "361",
      "species": "59",
      "source_groups": "335",
      "method": "source_cluster",
      "estimate": "0.30946223898288755",
      "ci_low": "0.23368183095623263",
      "ci_high": "0.3216250625222371",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5893333333333334",
      "ci_low": "0.5376344086021505",
      "ci_high": "0.6413043478260869",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_answer_coverage",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.9893333333333333",
      "ci_low": "0.9778393351800554",
      "ci_high": "0.9974620895714195",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "species_attempted_accuracy",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5956873315363881",
      "ci_low": "0.5437113413209914",
      "ci_high": "0.6472222222222223",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "wrong_species_rate_all_events",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.34908136482939633",
      "ci_high": "0.4520547945205479",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "any_rank_conflict_rate",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.42933333333333334",
      "ci_low": "0.3775201932053416",
      "ci_high": "0.48128591136648075",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "raw_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "74.16",
      "ci_low": "70.54429843622248",
      "ci_high": "77.67205062205062",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "adjusted_tis_pct",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "31.226666666666667",
      "ci_low": "22.81810636856369",
      "ci_high": "39.522505350454786",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.3341092241092241",
      "ci_low": "0.28877395047234",
      "ci_high": "0.3394712910434869",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "easy",
      "metric": "reference_species_macro_f1",
      "events": "375",
      "species": "78",
      "source_groups": "357",
      "method": "source_cluster",
      "estimate": "0.5568487068487068",
      "ci_low": "0.48128991745389993",
      "ci_high": "0.5657854850724782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.1206896551724138",
      "ci_low": "0.07262569832402235",
      "ci_high": "0.17318435754189945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_answer_coverage",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.5402298850574713",
      "ci_low": "0.4634146341463415",
      "ci_high": "0.6170243161094224",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "species_attempted_accuracy",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.22340425531914893",
      "ci_low": "0.13829787234042554",
      "ci_high": "0.3113207547169811",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "wrong_species_rate_all_events",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.41954022988505746",
      "ci_low": "0.3465909090909091",
      "ci_high": "0.49404761904761907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "any_rank_conflict_rate",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.4367816091954023",
      "ci_low": "0.3619631901840491",
      "ci_high": "0.5126590310351474",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "raw_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "23.218390804597703",
      "ci_low": "18.125",
      "ci_high": "28.54670425889605",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "adjusted_tis_pct",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "-20.45977011494253",
      "ci_low": "-28.80250578388677",
      "ci_high": "-11.904511278195493",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.07115812536373284",
      "ci_low": "0.03657070538495772",
      "ci_high": "0.08156063303259564",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "hard",
      "metric": "reference_species_macro_f1",
      "events": "174",
      "species": "77",
      "source_groups": "169",
      "method": "source_cluster",
      "estimate": "0.09888207031064172",
      "ci_low": "0.05081903215831787",
      "ci_high": "0.11333750304529523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.32083333333333336",
      "ci_low": "0.2616033755274262",
      "ci_high": "0.38257658102766795",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_answer_coverage",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.8916666666666667",
      "ci_low": "0.8508754244482173",
      "ci_high": "0.9297520661157025",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "species_attempted_accuracy",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3598130841121495",
      "ci_low": "0.2949293985584308",
      "ci_high": "0.42660550458715596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "wrong_species_rate_all_events",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5708333333333333",
      "ci_low": "0.5061728395061729",
      "ci_high": "0.6347860593512767",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "any_rank_conflict_rate",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.5958333333333333",
      "ci_low": "0.5321078320603787",
      "ci_high": "0.6600790513833992",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "raw_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "51.833333333333336",
      "ci_low": "46.94684119960669",
      "ci_high": "56.78862720966659",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "adjusted_tis_pct",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "-7.75",
      "ci_low": "-17.3556479086493",
      "ci_high": "2.2326513590422907",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.2151797535797536",
      "ci_low": "0.15615385336885335",
      "ci_high": "0.21379193192772142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "medium",
      "metric": "reference_species_macro_f1",
      "events": "240",
      "species": "79",
      "source_groups": "234",
      "method": "source_cluster",
      "estimate": "0.3404742936388506",
      "ci_low": "0.24707888191274266",
      "ci_high": "0.33827837330335664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty",
      "stratum": "unset",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.455",
      "ci_low": "0.40540540540540543",
      "ci_high": "0.50495111078272",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_answer_coverage",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.875",
      "ci_low": "0.8414322250639387",
      "ci_high": "0.9066994942608523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "species_attempted_accuracy",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.52",
      "ci_low": "0.4668587896253602",
      "ci_high": "0.5738185781760929",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "wrong_species_rate_all_events",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.42",
      "ci_low": "0.36939313984168864",
      "ci_high": "0.469334693877551",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "any_rank_conflict_rate",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4525",
      "ci_low": "0.40245962672855623",
      "ci_high": "0.5024878731343284",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "raw_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "60.099999999999994",
      "ci_low": "55.97328909952606",
      "ci_high": "64.11488330696201",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "adjusted_tis_pct",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "14.850000000000001",
      "ci_low": "6.828572780154917",
      "ci_high": "22.700588957941896",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.2042805257464443",
      "ci_low": "0.1858802338328159",
      "ci_high": "0.21486251008167706",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Aves",
      "metric": "reference_species_macro_f1",
      "events": "400",
      "species": "40",
      "source_groups": "367",
      "method": "source_cluster",
      "estimate": "0.4953802749351276",
      "ci_low": "0.4507595670445786",
      "ci_high": "0.5210415869480669",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.35384615384615387",
      "ci_low": "0.3036649214659686",
      "ci_high": "0.40601659535579276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_answer_coverage",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.8461538461538461",
      "ci_low": "0.8068459657701712",
      "ci_high": "0.883419689119171",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "species_attempted_accuracy",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.41818181818181815",
      "ci_low": "0.36079316348973606",
      "ci_high": "0.47648955418568933",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "wrong_species_rate_all_events",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.49230769230769234",
      "ci_low": "0.43902439024390244",
      "ci_high": "0.5467989648936833",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "any_rank_conflict_rate",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.5102564102564102",
      "ci_low": "0.4565189520624303",
      "ci_high": "0.5646446199083737",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "raw_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "52.17948717948718",
      "ci_low": "47.94665525672372",
      "ci_high": "56.480025316455695",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "adjusted_tis_pct",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "1.153846153846154",
      "ci_low": "-6.855670103092784",
      "ci_high": "9.30853267897026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.1860547887327449",
      "ci_low": "0.15939678648592506",
      "ci_high": "0.20091256108108368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class",
      "stratum": "Mammalia",
      "metric": "reference_species_macro_f1",
      "events": "390",
      "species": "39",
      "source_groups": "354",
      "method": "source_cluster",
      "estimate": "0.38165084868255367",
      "ci_low": "0.32696776715061543",
      "ci_high": "0.4121283304227359",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5780730897009967",
      "ci_low": "0.5218818620992534",
      "ci_high": "0.6330291729918817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8471760797342193",
      "ci_low": "0.8049623767514271",
      "ci_high": "0.887720193340494",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6823529411764706",
      "ci_low": "0.6245349599054253",
      "ci_high": "0.7381822083495903",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2691029900332226",
      "ci_low": "0.21981339094342192",
      "ci_high": "0.32094594594594594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.32558139534883723",
      "ci_low": "0.2743055555555556",
      "ci_high": "0.37888362209872506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "65.98006644518273",
      "ci_low": "61.170740137642866",
      "ci_high": "70.71428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "33.421926910299",
      "ci_low": "24.934640522875817",
      "ci_high": "41.61491216836354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.3747442942594927",
      "ci_low": "0.31435163568859653",
      "ci_high": "0.3865324167486628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Belgium",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6183280855281629",
      "ci_low": "0.5186801988861843",
      "ci_high": "0.6377784876352937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2074074074074074",
      "ci_low": "0.1544387997019576",
      "ci_high": "0.2647058823529412",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_answer_coverage",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.8518518518518519",
      "ci_low": "0.8042637010676156",
      "ci_high": "0.8958368288590604",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "species_attempted_accuracy",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.24347826086956523",
      "ci_low": "0.18292096622889306",
      "ci_high": "0.30932440905981606",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "wrong_species_rate_all_events",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6444444444444445",
      "ci_low": "0.5789473684210527",
      "ci_high": "0.705301221828636",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "any_rank_conflict_rate",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.6444444444444445",
      "ci_low": "0.5789473684210527",
      "ci_high": "0.705301221828636",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "raw_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "41.48148148148148",
      "ci_low": "36.69982283464567",
      "ci_high": "46.34761357101782",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "adjusted_tis_pct",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "-22.962962962962962",
      "ci_low": "-32.015276876487135",
      "ci_high": "-13.518055555555563",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.07860060442246364",
      "ci_low": "0.06058258164397796",
      "ci_high": "0.08958801569651421",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Ecuador",
      "metric": "reference_species_macro_f1",
      "events": "270",
      "species": "27",
      "source_groups": "226",
      "method": "source_cluster",
      "estimate": "0.2212461457817495",
      "ci_low": "0.1705287483311972",
      "ci_high": "0.25217367381241035",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.410958904109589",
      "ci_low": "0.3417024145115823",
      "ci_high": "0.4772727272727273",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_answer_coverage",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.8904109589041096",
      "ci_low": "0.8471615720524017",
      "ci_high": "0.930739421868454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "species_attempted_accuracy",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.46153846153846156",
      "ci_low": "0.38678758996304224",
      "ci_high": "0.532608695652174",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "wrong_species_rate_all_events",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4794520547945205",
      "ci_low": "0.4131455399061033",
      "ci_high": "0.5485237188152882",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "any_rank_conflict_rate",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.4931506849315068",
      "ci_low": "0.4266666666666667",
      "ci_high": "0.5627906976744186",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "raw_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "60.86757990867579",
      "ci_low": "55.575126422250314",
      "ci_high": "66.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "adjusted_tis_pct",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "11.552511415525116",
      "ci_low": "1.0253623188405798",
      "ci_high": "21.75008886255924",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.26716329707618997",
      "ci_low": "0.20561165159729788",
      "ci_high": "0.2835337488332451",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "country",
      "stratum": "Luxembourg",
      "metric": "reference_species_macro_f1",
      "events": "219",
      "species": "33",
      "source_groups": "197",
      "method": "source_cluster",
      "estimate": "0.47765559174227906",
      "ci_low": "0.3676087104315326",
      "ci_high": "0.5069239751867107",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.5821917808219178",
      "ci_low": "0.5253595740008784",
      "ci_high": "0.6382270713818894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_answer_coverage",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.8424657534246576",
      "ci_low": "0.8",
      "ci_high": "0.8841085744161729",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "species_attempted_accuracy",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6910569105691057",
      "ci_low": "0.6322314049586777",
      "ci_high": "0.7473047985687357",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "wrong_species_rate_all_events",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.2602739726027397",
      "ci_low": "0.21111014994232988",
      "ci_high": "0.312280701754386",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "any_rank_conflict_rate",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.3184931506849315",
      "ci_low": "0.2664359861591695",
      "ci_high": "0.3727598566308244",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "raw_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "66.26712328767124",
      "ci_low": "61.3604710095959",
      "ci_high": "71.07269338844894",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "adjusted_tis_pct",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "34.417808219178085",
      "ci_low": "25.910468779535",
      "ci_high": "42.64826084344586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.38590240723266706",
      "ci_low": "0.3174225095315988",
      "ci_high": "0.39435467263546026",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "10",
      "metric": "reference_species_macro_f1",
      "events": "292",
      "species": "41",
      "source_groups": "292",
      "method": "source_cluster",
      "estimate": "0.6212087531062446",
      "ci_low": "0.5109728202215981",
      "ci_high": "0.6348148388765946",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4052631578947368",
      "ci_low": "0.3315500755638224",
      "ci_high": "0.4759672619047618",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_answer_coverage",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.8736842105263158",
      "ci_low": "0.8163237727523442",
      "ci_high": "0.9257149929278642",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "species_attempted_accuracy",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.463855421686747",
      "ci_low": "0.38460264900662255",
      "ci_high": "0.5403726708074534",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "wrong_species_rate_all_events",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.46842105263157896",
      "ci_low": "0.39654917363803305",
      "ci_high": "0.5454545454545454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "any_rank_conflict_rate",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4842105263157895",
      "ci_low": "0.4111111111111111",
      "ci_high": "0.5612244897959183",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "raw_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "59.21052631578948",
      "ci_low": "53.238595310036786",
      "ci_high": "64.79606991311377",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "adjusted_tis_pct",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "10.789473684210527",
      "ci_low": "-0.621126785275643",
      "ci_high": "21.571482958665698",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.23327391256692753",
      "ci_low": "0.16734852808911027",
      "ci_high": "0.24091852449307474",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "2",
      "metric": "reference_species_macro_f1",
      "events": "190",
      "species": "41",
      "source_groups": "171",
      "method": "source_cluster",
      "estimate": "0.4153413565216027",
      "ci_low": "0.29796201342695244",
      "ci_high": "0.42895249482913306",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.24719101123595505",
      "ci_low": "0.14633461222252672",
      "ci_high": "0.3561683199496142",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_answer_coverage",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.8764044943820225",
      "ci_low": "0.8048780487804879",
      "ci_high": "0.9405940594059405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "species_attempted_accuracy",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.28205128205128205",
      "ci_low": "0.16901408450704225",
      "ci_high": "0.40229885057471265",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "wrong_species_rate_all_events",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6292134831460674",
      "ci_low": "0.5164746543778802",
      "ci_high": "0.735632183908046",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "any_rank_conflict_rate",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.6292134831460674",
      "ci_low": "0.5164746543778802",
      "ci_high": "0.735632183908046",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "raw_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "47.865168539325836",
      "ci_low": "39.404582616179",
      "ci_high": "56.78607142857143",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "adjusted_tis_pct",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "-15.056179775280897",
      "ci_low": "-31.29426470588235",
      "ci_high": "2.261904761904762",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.07764820213799806",
      "ci_low": "0.042996289424860855",
      "ci_high": "0.09657312925170067",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "3",
      "metric": "reference_species_macro_f1",
      "events": "89",
      "species": "22",
      "source_groups": "75",
      "method": "source_cluster",
      "estimate": "0.17294372294372293",
      "ci_low": "0.09576446280991736",
      "ci_high": "0.21509469696969696",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.25",
      "ci_low": "0.0",
      "ci_high": "0.5555555555555556",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_answer_coverage",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.8333333333333334",
      "ci_low": "0.625",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "species_attempted_accuracy",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.3",
      "ci_low": "0.0",
      "ci_high": "0.625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "wrong_species_rate_all_events",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5833333333333334",
      "ci_low": "0.3",
      "ci_high": "0.8571428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "any_rank_conflict_rate",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.5833333333333334",
      "ci_low": "0.3",
      "ci_high": "0.8571428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "raw_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "40.83333333333333",
      "ci_low": "18.88541666666667",
      "ci_high": "67.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "adjusted_tis_pct",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "-17.5",
      "ci_low": "-58.3392857142857",
      "ci_high": "31.11111111111111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.125",
      "ci_low": "0.0",
      "ci_high": "0.125",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "4",
      "metric": "reference_species_macro_f1",
      "events": "12",
      "species": "9",
      "source_groups": "11",
      "method": "source_cluster",
      "estimate": "0.2222222222222222",
      "ci_low": "0.0",
      "ci_high": "0.2222222222222222",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.22388059701492538",
      "ci_low": "0.1650485436893204",
      "ci_high": "0.2857142857142857",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_answer_coverage",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.8656716417910447",
      "ci_low": "0.8153846153846154",
      "ci_high": "0.9112178306707794",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "species_attempted_accuracy",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.25862068965517243",
      "ci_low": "0.19230769230769232",
      "ci_high": "0.32934131736526945",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "wrong_species_rate_all_events",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6417910447761194",
      "ci_low": "0.5714285714285714",
      "ci_high": "0.7081389553429027",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "any_rank_conflict_rate",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.6417910447761194",
      "ci_low": "0.5714285714285714",
      "ci_high": "0.7081389553429027",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "raw_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "43.23383084577115",
      "ci_low": "37.94724880382775",
      "ci_high": "48.54169283919598",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "adjusted_tis_pct",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "-20.945273631840795",
      "ci_low": "-31.01390299148637",
      "ci_high": "-10.531902045431185",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.10220991377408489",
      "ci_low": "0.0706708613197345",
      "ci_high": "0.11347885674774923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "5",
      "metric": "reference_species_macro_f1",
      "events": "201",
      "species": "27",
      "source_groups": "188",
      "method": "source_cluster",
      "estimate": "0.24984645589220753",
      "ci_low": "0.17275099433712876",
      "ci_high": "0.2773927609389426",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "55.0",
      "ci_low": "10.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "5.0",
      "ci_low": "-90.0",
      "ci_high": "100.0",
      "valid_resamples": "8667",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "7",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "1",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.6666666666666666",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "-80.0",
      "ci_high": "100.0",
      "valid_resamples": "8628",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "8",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_answer_coverage",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "1.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "species_attempted_accuracy",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "wrong_species_rate_all_events",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "any_rank_conflict_rate",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "1.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "raw_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "20.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "adjusted_tis_pct",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "10.0",
      "ci_low": "-80.0",
      "ci_high": "100.0",
      "valid_resamples": "8687",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.3333333333333333",
      "ci_low": "0.0",
      "ci_high": "0.3333333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "frame_count",
      "stratum": "9",
      "metric": "reference_species_macro_f1",
      "events": "2",
      "species": "2",
      "source_groups": "2",
      "method": "source_cluster",
      "estimate": "0.5",
      "ci_low": "0.0",
      "ci_high": "0.5",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.2714285714285714",
      "ci_low": "0.15554597701149425",
      "ci_high": "0.3918918918918919",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_answer_coverage",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.8857142857142857",
      "ci_low": "0.8051948051948052",
      "ci_high": "0.9552238805970149",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "species_attempted_accuracy",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3064516129032258",
      "ci_low": "0.17855158730158732",
      "ci_high": "0.43859649122807015",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "wrong_species_rate_all_events",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6142857142857143",
      "ci_low": "0.4918032786885246",
      "ci_high": "0.7391304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "any_rank_conflict_rate",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.6142857142857143",
      "ci_low": "0.4918032786885246",
      "ci_high": "0.7391304347826086",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "raw_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "50.42857142857143",
      "ci_low": "41.14251152073732",
      "ci_high": "59.63652597402598",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "adjusted_tis_pct",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "-11.0",
      "ci_low": "-29.51229346970889",
      "ci_high": "8.197282730743312",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.09838243597834391",
      "ci_low": "0.06452431927975406",
      "ci_high": "0.12496990990916822",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "1290a78f-098c-4336-84e4-74de27a658f1",
      "metric": "reference_species_macro_f1",
      "events": "70",
      "species": "7",
      "source_groups": "61",
      "method": "source_cluster",
      "estimate": "0.3232565753574157",
      "ci_low": "0.21200847763347763",
      "ci_high": "0.41061541827298126",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.2682926829268293",
      "ci_high": "0.5277908805031445",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_answer_coverage",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.7777777777777778",
      "ci_low": "0.65",
      "ci_high": "0.8947368421052632",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "species_attempted_accuracy",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5142857142857142",
      "ci_low": "0.36363636363636365",
      "ci_high": "0.6585365853658537",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "wrong_species_rate_all_events",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.37777777777777777",
      "ci_low": "0.25",
      "ci_high": "0.5116279069767442",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "any_rank_conflict_rate",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.4",
      "ci_low": "0.2727272727272727",
      "ci_high": "0.5365853658536586",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "raw_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "55.33333333333333",
      "ci_low": "43.61659574468086",
      "ci_high": "66.66836734693875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "adjusted_tis_pct",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "15.333333333333334",
      "ci_low": "-3.043780193236715",
      "ci_high": "33.19205927051671",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.3279799565513851",
      "ci_low": "0.14681234335839602",
      "ci_high": "0.34365740740740736",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "273ee7a0-4b59-4350-b220-3282b533ecde",
      "metric": "reference_species_macro_f1",
      "events": "45",
      "species": "13",
      "source_groups": "40",
      "method": "source_cluster",
      "estimate": "0.5298137759676222",
      "ci_low": "0.23715840080971665",
      "ci_high": "0.5551388888888887",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.2088888888888889",
      "ci_low": "0.15085889959432047",
      "ci_high": "0.2727272727272727",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_answer_coverage",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.8533333333333334",
      "ci_low": "0.8",
      "ci_high": "0.9009445637960022",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "species_attempted_accuracy",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.24479166666666666",
      "ci_low": "0.17815715833182885",
      "ci_high": "0.3173076923076923",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "wrong_species_rate_all_events",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6444444444444445",
      "ci_low": "0.5731695266557408",
      "ci_high": "0.7131163426102427",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "any_rank_conflict_rate",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.6444444444444445",
      "ci_low": "0.5731695266557408",
      "ci_high": "0.7131163426102427",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "raw_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "41.68888888888888",
      "ci_low": "36.38442028985507",
      "ci_high": "47.17086156278229",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "adjusted_tis_pct",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "-22.755555555555556",
      "ci_low": "-32.76787559031877",
      "ci_high": "-12.02641747683427",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.0801458123826545",
      "ci_low": "0.05808334547305136",
      "ci_high": "0.09185670200264393",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "68f49e46-9480-41dc-a51c-f600bb9a5f60",
      "metric": "reference_species_macro_f1",
      "events": "225",
      "species": "27",
      "source_groups": "182",
      "method": "source_cluster",
      "estimate": "0.21965889319690493",
      "ci_low": "0.1591913912965111",
      "ci_high": "0.25175540548872777",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.28169014084507044",
      "ci_low": "0.1780821917808219",
      "ci_high": "0.390625",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_answer_coverage",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.8873239436619719",
      "ci_low": "0.8088235294117647",
      "ci_high": "0.9558983375959077",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "species_attempted_accuracy",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.31746031746031744",
      "ci_low": "0.2033898305084746",
      "ci_high": "0.4383561643835616",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "wrong_species_rate_all_events",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.49019607843137253",
      "ci_high": "0.71875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "any_rank_conflict_rate",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.6056338028169014",
      "ci_low": "0.49019607843137253",
      "ci_high": "0.71875",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "raw_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "46.90140845070423",
      "ci_low": "37.94498804087846",
      "ci_high": "55.91556481747629",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "adjusted_tis_pct",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "-13.661971830985916",
      "ci_low": "-31.0356477166822",
      "ci_high": "4.69796650717702",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.20703463203463202",
      "ci_low": "0.11009920634920636",
      "ci_high": "0.22435749540863179",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "None",
      "metric": "reference_species_macro_f1",
      "events": "71",
      "species": "27",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.33738977072310405",
      "ci_low": "0.17942092886537334",
      "ci_high": "0.36561962214739996",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.5780730897009967",
      "ci_low": "0.5218818620992534",
      "ci_high": "0.6330291729918817",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_answer_coverage",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.8471760797342193",
      "ci_low": "0.8049623767514271",
      "ci_high": "0.887720193340494",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "species_attempted_accuracy",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6823529411764706",
      "ci_low": "0.6245349599054253",
      "ci_high": "0.7381822083495903",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "wrong_species_rate_all_events",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.2691029900332226",
      "ci_low": "0.21981339094342192",
      "ci_high": "0.32094594594594594",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "any_rank_conflict_rate",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.32558139534883723",
      "ci_low": "0.2743055555555556",
      "ci_high": "0.37888362209872506",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "raw_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "65.98006644518273",
      "ci_low": "61.170740137642866",
      "ci_high": "70.71428571428571",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "adjusted_tis_pct",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "33.421926910299",
      "ci_low": "24.934640522875817",
      "ci_high": "41.61491216836354",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.3747442942594927",
      "ci_low": "0.31435163568859653",
      "ci_high": "0.3865324167486628",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "c9cbc586-660e-4d89-ba14-0000c5770de1",
      "metric": "reference_species_macro_f1",
      "events": "301",
      "species": "40",
      "source_groups": "298",
      "method": "source_cluster",
      "estimate": "0.6183280855281629",
      "ci_low": "0.5186801988861843",
      "ci_high": "0.6377784876352937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5384615384615384",
      "ci_low": "0.4177215189873418",
      "ci_high": "0.6565694255992763",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_answer_coverage",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.9358974358974359",
      "ci_low": "0.8765432098765432",
      "ci_high": "0.9863013698630136",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "species_attempted_accuracy",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.5753424657534246",
      "ci_low": "0.45",
      "ci_high": "0.6935483870967742",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "wrong_species_rate_all_events",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.3974358974358974",
      "ci_low": "0.2857142857142857",
      "ci_high": "0.5172413793103449",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "any_rank_conflict_rate",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.4230769230769231",
      "ci_low": "0.30985915492957744",
      "ci_high": "0.5454545454545454",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "raw_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "74.35897435897436",
      "ci_low": "66.58528866934239",
      "ci_high": "81.7074912891986",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "adjusted_tis_pct",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "32.05128205128205",
      "ci_low": "14.507042253521128",
      "ci_high": "48.98876404494382",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.36635243841126197",
      "ci_low": "0.2716320831945832",
      "ci_high": "0.3977283133533133",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "dataset_key",
      "stratum": "fc3f505a-05d8-4b3e-908c-8880fc9899f7",
      "metric": "reference_species_macro_f1",
      "events": "78",
      "species": "16",
      "source_groups": "70",
      "method": "source_cluster",
      "estimate": "0.549528657616893",
      "ci_low": "0.40744812479187476",
      "ci_high": "0.59659247002997",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5092592592592593",
      "ci_low": "0.45364238410596025",
      "ci_high": "0.5641891891891891",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_answer_coverage",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.8734567901234568",
      "ci_low": "0.8368564371968766",
      "ci_high": "0.9093569459903535",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "species_attempted_accuracy",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5830388692579506",
      "ci_low": "0.5243055555555556",
      "ci_high": "0.6411149825783972",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "wrong_species_rate_all_events",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.36419753086419754",
      "ci_low": "0.3111111111111111",
      "ci_high": "0.4171789298338276",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "any_rank_conflict_rate",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.404320987654321",
      "ci_low": "0.3501577287066246",
      "ci_high": "0.4591564188757925",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "raw_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "64.62962962962963",
      "ci_low": "60.12617304525524",
      "ci_high": "68.94260466119982",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "adjusted_tis_pct",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "24.19753086419753",
      "ci_low": "15.572243672677287",
      "ci_high": "32.477907854984885",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.2564240868637153",
      "ci_low": "0.2322738089112476",
      "ci_high": "0.2707035669158167",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | day",
      "metric": "reference_species_macro_f1",
      "events": "324",
      "species": "38",
      "source_groups": "301",
      "method": "source_cluster",
      "estimate": "0.5061001714415434",
      "ci_low": "0.4584351491669362",
      "ci_high": "0.5342833557549014",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2236842105263158",
      "ci_low": "0.1323529411764706",
      "ci_high": "0.32142857142857145",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.881578947368421",
      "ci_low": "0.8028169014084507",
      "ci_high": "0.95",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2537313432835821",
      "ci_low": "0.1506849315068493",
      "ci_high": "0.36363636363636365",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6578947368421053",
      "ci_low": "0.5443037974683544",
      "ci_high": "0.7666666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.6578947368421053",
      "ci_low": "0.5443037974683544",
      "ci_high": "0.7666666666666667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "40.78947368421052",
      "ci_low": "32.49932795698925",
      "ci_high": "49.48717948717949",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "-25.0",
      "ci_low": "-41.46693181818182",
      "ci_high": "-7.71345755693583",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.10329004329004329",
      "ci_low": "0.05166590909090908",
      "ci_high": "0.11227152406417111",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Aves | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "20",
      "source_groups": "71",
      "method": "source_cluster",
      "estimate": "0.2582251082251082",
      "ci_low": "0.12916477272727273",
      "ci_high": "0.28067881016042784",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.42857142857142855",
      "ci_low": "0.3220290278589519",
      "ci_high": "0.5339805825242718",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_answer_coverage",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.8761904761904762",
      "ci_low": "0.81",
      "ci_high": "0.9363636363636364",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "species_attempted_accuracy",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4891304347826087",
      "ci_low": "0.37142857142857144",
      "ci_high": "0.6049382716049383",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "wrong_species_rate_all_events",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.44761904761904764",
      "ci_low": "0.339995145631068",
      "ci_high": "0.5575221238938053",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "any_rank_conflict_rate",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.44761904761904764",
      "ci_low": "0.339995145631068",
      "ci_high": "0.5575221238938053",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "raw_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "51.96246444534742",
      "ci_high": "67.89486842105262",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "adjusted_tis_pct",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "15.238095238095237",
      "ci_low": "-1.1768207282913163",
      "ci_high": "31.666666666666664",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.2921770224254696",
      "ci_low": "0.1935114689109603",
      "ci_high": "0.31441578346465565",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | day",
      "metric": "reference_species_macro_f1",
      "events": "105",
      "species": "24",
      "source_groups": "94",
      "method": "source_cluster",
      "estimate": "0.4260914910371432",
      "ci_low": "0.2822042254951504",
      "ci_high": "0.4585230175526228",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3263157894736842",
      "ci_low": "0.2690909090909091",
      "ci_high": "0.3843810096153846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_answer_coverage",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.8350877192982457",
      "ci_low": "0.7873134328358209",
      "ci_high": "0.8803995738841542",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3907563025210084",
      "ci_low": "0.3252032520325203",
      "ci_high": "0.45689655172413796",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5087719298245614",
      "ci_low": "0.44755244755244755",
      "ci_high": "0.5700008143322475",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.5333333333333333",
      "ci_low": "0.47183098591549294",
      "ci_high": "0.5947040974258715",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "raw_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "49.29824561403508",
      "ci_low": "44.412785220849905",
      "ci_high": "54.20289855072464",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "-4.035087719298246",
      "ci_low": "-12.968225014773456",
      "ci_high": "5.130126709915689",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.17004896020114546",
      "ci_low": "0.13144020489747812",
      "ci_high": "0.1843991466197147",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "class_time",
      "stratum": "Mammalia | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "285",
      "species": "39",
      "source_groups": "264",
      "method": "source_cluster",
      "estimate": "0.3357376906535436",
      "ci_low": "0.25951014813091833",
      "ci_high": "0.3640701099927701",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6857142857142857",
      "ci_low": "0.6234309623430963",
      "ci_high": "0.7449412874816523",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_answer_coverage",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.9918367346938776",
      "ci_low": "0.9789029535864979",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "species_attempted_accuracy",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.691358024691358",
      "ci_low": "0.6290322580645161",
      "ci_high": "0.7510377820963986",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "wrong_species_rate_all_events",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.30612244897959184",
      "ci_low": "0.24686192468619247",
      "ci_high": "0.368",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "any_rank_conflict_rate",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.34285714285714286",
      "ci_low": "0.2813658507604563",
      "ci_high": "0.40613326149425283",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "raw_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "81.6734693877551",
      "ci_low": "77.64257782311664",
      "ci_high": "85.49025178253119",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "adjusted_tis_pct",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "47.38775510204082",
      "ci_low": "37.363931187880105",
      "ci_high": "56.887989857076995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.4293459552495697",
      "ci_low": "0.3572486670288907",
      "ci_high": "0.43041251452611723",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | day",
      "metric": "reference_species_macro_f1",
      "events": "245",
      "species": "59",
      "source_groups": "231",
      "method": "source_cluster",
      "estimate": "0.6039951573849879",
      "ci_low": "0.5025701587016598",
      "ci_high": "0.6054955712825039",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4076923076923077",
      "ci_low": "0.32231404958677684",
      "ci_high": "0.4957282913165266",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_answer_coverage",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.9846153846153847",
      "ci_low": "0.96",
      "ci_high": "1.0",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4140625",
      "ci_low": "0.3274296057924377",
      "ci_high": "0.5035460992907801",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5769230769230769",
      "ci_low": "0.48836751510712323",
      "ci_high": "0.6635535259133389",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.5923076923076923",
      "ci_low": "0.5038744896147701",
      "ci_high": "0.6779661016949152",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "raw_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "60.0",
      "ci_low": "53.54608317214701",
      "ci_high": "66.49644160583942",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.7692307692307693",
      "ci_low": "-13.520114754098362",
      "ci_high": "15.397321428571422",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.2205654761904762",
      "ci_low": "0.15291540230602732",
      "ci_high": "0.229375744047619",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "easy | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "130",
      "species": "44",
      "source_groups": "127",
      "method": "source_cluster",
      "estimate": "0.4010281385281385",
      "ci_low": "0.27802800419277696",
      "ci_high": "0.4170468073593073",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.06578947368421052",
      "ci_low": "0.013513513513513514",
      "ci_high": "0.12698412698412698",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_answer_coverage",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.4605263157894737",
      "ci_low": "0.34882840028188866",
      "ci_high": "0.5731747967479673",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "species_attempted_accuracy",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.14285714285714285",
      "ci_low": "0.03125",
      "ci_high": "0.2702702702702703",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "wrong_species_rate_all_events",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.39473684210526316",
      "ci_low": "0.287496408045977",
      "ci_high": "0.5061728395061729",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "any_rank_conflict_rate",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.42105263157894735",
      "ci_low": "0.3125",
      "ci_high": "0.5316455696202531",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "raw_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "17.105263157894736",
      "ci_low": "10.833333333333332",
      "ci_high": "24.050632911392405",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "adjusted_tis_pct",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "-25.0",
      "ci_low": "-35.21790890269151",
      "ci_high": "-14.931169997754328",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.048205128205128206",
      "ci_low": "0.010256410256410256",
      "ci_high": "0.05846153846153846",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | day",
      "metric": "reference_species_macro_f1",
      "events": "76",
      "species": "46",
      "source_groups": "76",
      "method": "source_cluster",
      "estimate": "0.06811594202898551",
      "ci_low": "0.014492753623188404",
      "ci_high": "0.08260869565217391",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.16326530612244897",
      "ci_low": "0.09090909090909091",
      "ci_high": "0.24691721132897596",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_answer_coverage",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.6020408163265306",
      "ci_low": "0.5",
      "ci_high": "0.7040816326530612",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.2711864406779661",
      "ci_low": "0.15623263888888891",
      "ci_high": "0.39285714285714285",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4387755102040816",
      "ci_low": "0.33962264150943394",
      "ci_high": "0.5421686746987951",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.4489795918367347",
      "ci_low": "0.3474553969669938",
      "ci_high": "0.5505732630130702",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "raw_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "27.95918367346939",
      "ci_low": "20.574630541871922",
      "ci_high": "35.94623346751006",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "-16.93877551020408",
      "ci_low": "-29.419552578361976",
      "ci_high": "-4.225273865414711",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.08222912353347137",
      "ci_low": "0.039195709684840124",
      "ci_high": "0.09879227053140097",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "hard | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "98",
      "species": "46",
      "source_groups": "93",
      "method": "source_cluster",
      "estimate": "0.12334368530020705",
      "ci_low": "0.058793564527260186",
      "ci_high": "0.14818840579710144",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.3364485981308411",
      "ci_low": "0.24770642201834864",
      "ci_high": "0.42609369733737773",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_answer_coverage",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.897196261682243",
      "ci_low": "0.8365384615384616",
      "ci_high": "0.9516206575682381",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "species_attempted_accuracy",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.375",
      "ci_low": "0.2792740414833438",
      "ci_high": "0.4717034131863472",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "wrong_species_rate_all_events",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5607476635514018",
      "ci_low": "0.46609963797926607",
      "ci_high": "0.6551724137931034",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "any_rank_conflict_rate",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.5794392523364486",
      "ci_low": "0.4891304347826087",
      "ci_high": "0.6728971962616822",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "raw_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "54.48598130841121",
      "ci_low": "47.5",
      "ci_high": "61.61926910299003",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "adjusted_tis_pct",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "-3.4579439252336446",
      "ci_low": "-17.403935185185183",
      "ci_high": "10.488236018723818",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.2309206349206349",
      "ci_low": "0.14162012987012987",
      "ci_high": "0.22733333333333333",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | day",
      "metric": "reference_species_macro_f1",
      "events": "107",
      "species": "51",
      "source_groups": "106",
      "method": "source_cluster",
      "estimate": "0.33958916900093367",
      "ci_low": "0.20826489686783803",
      "ci_high": "0.3343137254901961",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3082706766917293",
      "ci_low": "0.22880750605326877",
      "ci_high": "0.3944954128440367",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_answer_coverage",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.8872180451127819",
      "ci_low": "0.8309859154929577",
      "ci_high": "0.9396551724137931",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "species_attempted_accuracy",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3474576271186441",
      "ci_low": "0.2601626016260163",
      "ci_high": "0.44118309485956536",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "wrong_species_rate_all_events",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.5789473684210527",
      "ci_low": "0.4895067698259188",
      "ci_high": "0.6642335766423357",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "any_rank_conflict_rate",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.6090225563909775",
      "ci_low": "0.5220588235294118",
      "ci_high": "0.6936936936936937",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "raw_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "49.69924812030075",
      "ci_low": "43.09515977443609",
      "ci_high": "56.559999999999995",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "adjusted_tis_pct",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "-11.203007518796992",
      "ci_low": "-23.931623931623932",
      "ci_high": "2.307823129251699",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.19558103076621594",
      "ci_low": "0.12479618571938833",
      "ci_high": "0.19909069631528667",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "medium | night_ir",
      "metric": "reference_species_macro_f1",
      "events": "133",
      "species": "49",
      "source_groups": "129",
      "method": "source_cluster",
      "estimate": "0.3233074182053774",
      "ci_low": "0.20629573557694805",
      "ci_high": "0.3291091102354738",
      "valid_resamples": "10000",
      "interval_status": "ok"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_answer_coverage",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "species_attempted_accuracy",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "wrong_species_rate_all_events",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "any_rank_conflict_rate",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "0.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "raw_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "adjusted_tis_pct",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "100.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "6327",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    },
    {
      "configuration": "Qwen Instruct primary_repaired",
      "axis": "difficulty_time",
      "stratum": "unset | day",
      "metric": "reference_species_macro_f1",
      "events": "1",
      "species": "1",
      "source_groups": "1",
      "method": "source_cluster",
      "estimate": "1.0",
      "ci_low": "",
      "ci_high": "",
      "valid_resamples": "10000",
      "interval_status": "insufficient_support"
    }
  ],
  "unit_policy": {
    "table": "subgroup_intervals.csv",
    "metric_units": {
      "adjusted_tis_pct": "percent_of_five_point_maximum",
      "any_rank_conflict_rate": "proportion_0_to_1",
      "macro_f1": "f1_0_to_1",
      "raw_tis_pct": "percent_of_five_point_maximum",
      "reference_species_macro_f1": "f1_0_to_1",
      "species_accuracy": "proportion_0_to_1",
      "species_answer_coverage": "proportion_0_to_1",
      "species_attempted_accuracy": "proportion_0_to_1",
      "wrong_species_rate_all_events": "proportion_0_to_1"
    },
    "counts": {
      "events": "events",
      "species": "species",
      "source_groups": "source_groups",
      "valid_resamples": "bootstrap_resamples"
    },
    "interval": "ci_low and ci_high use the same units as estimate; pointwise 95% when interval_status is ok",
    "estimate": "unit selected by each row's metric_units entry"
  },
  "note": "Saved aggregate candidate analysis; intervals are pointwise and exploratory tests have separate Holm-adjusted p-values."
}
