{
  "schema_version": 1,
  "columns": [
    "family",
    "baseline",
    "comparison",
    "metric",
    "source_swap_p",
    "interpretation",
    "holm_p_specialist_3_tests"
  ],
  "rows": [
    {
      "family": "BioCLIP geography exploratory",
      "baseline": "BioCLIP 2.5 / primary_blind",
      "comparison": "BioCLIP 2.5 / operational_country_ablation",
      "metric": "species_accuracy",
      "source_swap_p": "9.999000099990002e-05",
      "interpretation": "exploratory; separate three-test specialist family; group label exchangeability assumed",
      "holm_p_specialist_3_tests": "0.00029997000299970003"
    },
    {
      "family": "SpeciesNet geography exploratory",
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "metric": "species_accuracy",
      "source_swap_p": "0.0326967303269673",
      "interpretation": "exploratory; separate three-test specialist family; group label exchangeability assumed",
      "holm_p_specialist_3_tests": "0.0326967303269673"
    },
    {
      "family": "SpeciesNet geography exploratory",
      "baseline": "SpeciesNet v4.0.3b / primary_blind",
      "comparison": "SpeciesNet v4.0.3b / operational_country_ablation",
      "metric": "adjusted_tis_pct",
      "source_swap_p": "9.999000099990002e-05",
      "interpretation": "exploratory; separate three-test specialist family; group label exchangeability assumed",
      "holm_p_specialist_3_tests": "0.00029997000299970003"
    }
  ],
  "unit_policy": {
    "table": "specialist_paired_tests.csv",
    "metric_units": {
      "adjusted_tis_pct": "percent_of_five_point_maximum",
      "species_accuracy": "proportion_0_to_1"
    },
    "counts": {},
    "interval": "ci_low and ci_high use the same units as estimate; pointwise 95% when interval_status is ok",
    "p_values": "probability_0_to_1; named Holm columns are adjusted p-values, not interval adjustments",
    "estimate": "not_applicable"
  },
  "note": "Saved aggregate candidate analysis; intervals are pointwise and exploratory tests have separate Holm-adjusted p-values."
}