{ "schema_version": 1, "confidence_definition": "One ordinal, self-reported label for the whole taxonomic response; not a calibrated species probability or independent rank confidence.", "unit_of_analysis": "one camera-trap event", "metric_definition": { "species_accuracy": "exact species correct / all events in confidence bucket, including abstentions and technical nonresponses", "adjusted_tis_mean": "arithmetic mean of already-saved adjusted TIS points per event in confidence bucket; five-point scale", "species_attempts": "events with a correct or wrong species answer", "withheld_or_technical": "events without a species answer; technical is a subset" }, "scope": "private exploratory descriptive V1 primary VLM views; no intervals or newly adopted scientific result", "cohorts": { "v1": { "categories": { "all": { "n_events": 790, "models": { "gemini-minimal": { "label": "Gemini 3.5 Flash-Lite - MINIMAL", "n_events": 790, "buckets": { "low": { "events": 134, "species_correct": 1, "species_attempts": 1, "withheld_or_technical": 133, "technical": 0, "species_accuracy": 0.007462686567164179, "adjusted_tis_mean": 0.026119402985074626 }, "medium": { "events": 217, "species_correct": 44, "species_attempts": 94, "withheld_or_technical": 123, "technical": 0, "species_accuracy": 0.20276497695852536, "adjusted_tis_mean": 0.6082949308755761 }, "high": { "events": 436, "species_correct": 313, "species_attempts": 387, "withheld_or_technical": 49, "technical": 0, "species_accuracy": 0.7178899082568807, "adjusted_tis_mean": 3.05848623853211 }, "missing": { "events": 3, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 3, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-low": { "label": "Gemini 3.5 Flash-Lite - LOW", "n_events": 790, "buckets": { "low": { "events": 152, "species_correct": 1, "species_attempts": 8, "withheld_or_technical": 144, "technical": 0, "species_accuracy": 0.006578947368421052, "adjusted_tis_mean": -0.3980263157894737 }, "medium": { "events": 242, "species_correct": 61, "species_attempts": 123, "withheld_or_technical": 119, "technical": 0, "species_accuracy": 0.25206611570247933, "adjusted_tis_mean": 0.6962809917355371 }, "high": { "events": 395, "species_correct": 306, "species_attempts": 365, "withheld_or_technical": 30, "technical": 0, "species_accuracy": 0.7746835443037975, "adjusted_tis_mean": 3.4151898734177215 }, "missing": { "events": 1, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 1, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium": { "label": "Gemini 3.5 Flash-Lite - MEDIUM", "n_events": 790, "buckets": { "low": { "events": 70, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 70, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.16428571428571428 }, "medium": { "events": 321, "species_correct": 102, "species_attempts": 199, "withheld_or_technical": 122, "technical": 0, "species_accuracy": 0.3177570093457944, "adjusted_tis_mean": 0.7305295950155763 }, "high": { "events": 397, "species_correct": 297, "species_attempts": 346, "withheld_or_technical": 51, "technical": 0, "species_accuracy": 0.7481108312342569, "adjusted_tis_mean": 3.370277078085642 }, "missing": { "events": 2, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 2, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-high": { "label": "Gemini 3.5 Flash-Lite - HIGH", "n_events": 790, "buckets": { "low": { "events": 49, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 49, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.061224489795918366 }, "medium": { "events": 287, "species_correct": 88, "species_attempts": 191, "withheld_or_technical": 96, "technical": 0, "species_accuracy": 0.30662020905923343, "adjusted_tis_mean": 0.3623693379790941 }, "high": { "events": 451, "species_correct": 318, "species_attempts": 387, "withheld_or_technical": 64, "technical": 0, "species_accuracy": 0.70509977827051, "adjusted_tis_mean": 3.0454545454545454 }, "missing": { "events": 3, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 3, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium37": { "label": "Gemini 3.7 Flash - MEDIUM", "n_events": 790, "buckets": { "low": { "events": 40, "species_correct": 0, "species_attempts": 2, "withheld_or_technical": 38, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.175 }, "medium": { "events": 112, "species_correct": 24, "species_attempts": 38, "withheld_or_technical": 74, "technical": 0, "species_accuracy": 0.21428571428571427, "adjusted_tis_mean": 1.0758928571428572 }, "high": { "events": 611, "species_correct": 416, "species_attempts": 489, "withheld_or_technical": 122, "technical": 0, "species_accuracy": 0.6808510638297872, "adjusted_tis_mean": 3.318330605564648 }, "missing": { "events": 27, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 27, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium38": { "label": "Gemini 3.8 Flash - MEDIUM", "n_events": 790, "buckets": { "low": { "events": 34, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 34, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.07352941176470588 }, "medium": { "events": 141, "species_correct": 42, "species_attempts": 61, "withheld_or_technical": 80, "technical": 0, "species_accuracy": 0.2978723404255319, "adjusted_tis_mean": 1.6914893617021276 }, "high": { "events": 591, "species_correct": 406, "species_attempts": 461, "withheld_or_technical": 130, "technical": 0, "species_accuracy": 0.6869712351945855, "adjusted_tis_mean": 3.44839255499154 }, "missing": { "events": 24, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 24, "technical": 1, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini38-low": { "label": "Gemini 3.8 Flash - LOW", "n_events": 790, "buckets": { "low": { "events": 54, "species_correct": 0, "species_attempts": 3, "withheld_or_technical": 51, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.1111111111111111 }, "medium": { "events": 133, "species_correct": 36, "species_attempts": 54, "withheld_or_technical": 79, "technical": 0, "species_accuracy": 0.2706766917293233, "adjusted_tis_mean": 1.6879699248120301 }, "high": { "events": 571, "species_correct": 414, "species_attempts": 470, "withheld_or_technical": 101, "technical": 0, "species_accuracy": 0.7250437828371279, "adjusted_tis_mean": 3.584938704028021 }, "missing": { "events": 32, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 32, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini38-high": { "label": "Gemini 3.8 Flash - HIGH", "n_events": 790, "buckets": { "low": { "events": 19, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 19, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.3684210526315789 }, "medium": { "events": 169, "species_correct": 73, "species_attempts": 116, "withheld_or_technical": 53, "technical": 0, "species_accuracy": 0.4319526627218935, "adjusted_tis_mean": 1.5236686390532543 }, "high": { "events": 580, "species_correct": 433, "species_attempts": 495, "withheld_or_technical": 85, "technical": 0, "species_accuracy": 0.746551724137931, "adjusted_tis_mean": 3.5801724137931035 }, "missing": { "events": 22, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 22, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "qwen-repaired": { "label": "Qwen3-VL-30B Instruct - repaired", "n_events": 790, "buckets": { "low": { "events": 121, "species_correct": 0, "species_attempts": 11, "withheld_or_technical": 110, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.4049586776859504 }, "medium": { "events": 222, "species_correct": 33, "species_attempts": 222, "withheld_or_technical": 0, "technical": 0, "species_accuracy": 0.14864864864864866, "adjusted_tis_mean": -2.2387387387387387 }, "high": { "events": 447, "species_correct": 287, "species_attempts": 447, "withheld_or_technical": 0, "technical": 0, "species_accuracy": 0.6420581655480985, "adjusted_tis_mean": 1.936241610738255 }, "missing": { "events": 0, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 0, "technical": 0, "species_accuracy": null, "adjusted_tis_mean": null } } } } }, "birds": { "n_events": 400, "models": { "gemini-minimal": { "label": "Gemini 3.5 Flash-Lite - MINIMAL", "n_events": 400, "buckets": { "low": { "events": 59, "species_correct": 1, "species_attempts": 1, "withheld_or_technical": 58, "technical": 0, "species_accuracy": 0.01694915254237288, "adjusted_tis_mean": -0.03389830508474576 }, "medium": { "events": 91, "species_correct": 25, "species_attempts": 45, "withheld_or_technical": 46, "technical": 0, "species_accuracy": 0.27472527472527475, "adjusted_tis_mean": 0.8131868131868132 }, "high": { "events": 248, "species_correct": 174, "species_attempts": 220, "withheld_or_technical": 28, "technical": 0, "species_accuracy": 0.7016129032258065, "adjusted_tis_mean": 2.933467741935484 }, "missing": { "events": 2, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 2, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-low": { "label": "Gemini 3.5 Flash-Lite - LOW", "n_events": 400, "buckets": { "low": { "events": 65, "species_correct": 0, "species_attempts": 6, "withheld_or_technical": 59, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.3384615384615385 }, "medium": { "events": 103, "species_correct": 27, "species_attempts": 54, "withheld_or_technical": 49, "technical": 0, "species_accuracy": 0.2621359223300971, "adjusted_tis_mean": 0.441747572815534 }, "high": { "events": 231, "species_correct": 176, "species_attempts": 213, "withheld_or_technical": 18, "technical": 0, "species_accuracy": 0.7619047619047619, "adjusted_tis_mean": 3.346320346320346 }, "missing": { "events": 1, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 1, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium": { "label": "Gemini 3.5 Flash-Lite - MEDIUM", "n_events": 400, "buckets": { "low": { "events": 27, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 27, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.3148148148148148 }, "medium": { "events": 146, "species_correct": 50, "species_attempts": 103, "withheld_or_technical": 43, "technical": 0, "species_accuracy": 0.3424657534246575, "adjusted_tis_mean": 0.5068493150684932 }, "high": { "events": 226, "species_correct": 170, "species_attempts": 204, "withheld_or_technical": 22, "technical": 0, "species_accuracy": 0.7522123893805309, "adjusted_tis_mean": 3.331858407079646 }, "missing": { "events": 1, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 1, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-high": { "label": "Gemini 3.5 Flash-Lite - HIGH", "n_events": 400, "buckets": { "low": { "events": 13, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 13, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.11538461538461539 }, "medium": { "events": 125, "species_correct": 36, "species_attempts": 90, "withheld_or_technical": 35, "technical": 0, "species_accuracy": 0.288, "adjusted_tis_mean": -0.084 }, "high": { "events": 261, "species_correct": 188, "species_attempts": 233, "withheld_or_technical": 28, "technical": 0, "species_accuracy": 0.7203065134099617, "adjusted_tis_mean": 3.0306513409961684 }, "missing": { "events": 1, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 1, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium37": { "label": "Gemini 3.7 Flash - MEDIUM", "n_events": 400, "buckets": { "low": { "events": 11, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 11, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.22727272727272727 }, "medium": { "events": 55, "species_correct": 8, "species_attempts": 17, "withheld_or_technical": 38, "technical": 0, "species_accuracy": 0.14545454545454545, "adjusted_tis_mean": 0.6 }, "high": { "events": 322, "species_correct": 242, "species_attempts": 280, "withheld_or_technical": 42, "technical": 0, "species_accuracy": 0.7515527950310559, "adjusted_tis_mean": 3.549689440993789 }, "missing": { "events": 12, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 12, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium38": { "label": "Gemini 3.8 Flash - MEDIUM", "n_events": 400, "buckets": { "low": { "events": 11, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 11, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.4090909090909091 }, "medium": { "events": 52, "species_correct": 13, "species_attempts": 19, "withheld_or_technical": 33, "technical": 0, "species_accuracy": 0.25, "adjusted_tis_mean": 1.5673076923076923 }, "high": { "events": 326, "species_correct": 251, "species_attempts": 285, "withheld_or_technical": 41, "technical": 0, "species_accuracy": 0.7699386503067485, "adjusted_tis_mean": 3.642638036809816 }, "missing": { "events": 11, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 11, "technical": 1, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini38-low": { "label": "Gemini 3.8 Flash - LOW", "n_events": 400, "buckets": { "low": { "events": 25, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 25, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.12 }, "medium": { "events": 48, "species_correct": 13, "species_attempts": 22, "withheld_or_technical": 26, "technical": 0, "species_accuracy": 0.2708333333333333, "adjusted_tis_mean": 1.6041666666666667 }, "high": { "events": 314, "species_correct": 251, "species_attempts": 284, "withheld_or_technical": 30, "technical": 0, "species_accuracy": 0.7993630573248408, "adjusted_tis_mean": 3.770700636942675 }, "missing": { "events": 13, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 13, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini38-high": { "label": "Gemini 3.8 Flash - HIGH", "n_events": 400, "buckets": { "low": { "events": 6, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 6, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -1.25 }, "medium": { "events": 59, "species_correct": 20, "species_attempts": 39, "withheld_or_technical": 20, "technical": 0, "species_accuracy": 0.3389830508474576, "adjusted_tis_mean": 1.0423728813559323 }, "high": { "events": 325, "species_correct": 260, "species_attempts": 300, "withheld_or_technical": 25, "technical": 0, "species_accuracy": 0.8, "adjusted_tis_mean": 3.7584615384615385 }, "missing": { "events": 10, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 10, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "qwen-repaired": { "label": "Qwen3-VL-30B Instruct - repaired", "n_events": 400, "buckets": { "low": { "events": 56, "species_correct": 0, "species_attempts": 6, "withheld_or_technical": 50, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.48214285714285715 }, "medium": { "events": 87, "species_correct": 13, "species_attempts": 87, "withheld_or_technical": 0, "technical": 0, "species_accuracy": 0.14942528735632185, "adjusted_tis_mean": -2.5 }, "high": { "events": 257, "species_correct": 169, "species_attempts": 257, "withheld_or_technical": 0, "technical": 0, "species_accuracy": 0.6575875486381323, "adjusted_tis_mean": 2.1070038910505837 }, "missing": { "events": 0, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 0, "technical": 0, "species_accuracy": null, "adjusted_tis_mean": null } } } } }, "mammals": { "n_events": 390, "models": { "gemini-minimal": { "label": "Gemini 3.5 Flash-Lite - MINIMAL", "n_events": 390, "buckets": { "low": { "events": 75, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 75, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.07333333333333333 }, "medium": { "events": 126, "species_correct": 19, "species_attempts": 49, "withheld_or_technical": 77, "technical": 0, "species_accuracy": 0.15079365079365079, "adjusted_tis_mean": 0.4603174603174603 }, "high": { "events": 188, "species_correct": 139, "species_attempts": 167, "withheld_or_technical": 21, "technical": 0, "species_accuracy": 0.7393617021276596, "adjusted_tis_mean": 3.223404255319149 }, "missing": { "events": 1, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 1, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-low": { "label": "Gemini 3.5 Flash-Lite - LOW", "n_events": 390, "buckets": { "low": { "events": 87, "species_correct": 1, "species_attempts": 2, "withheld_or_technical": 85, "technical": 0, "species_accuracy": 0.011494252873563218, "adjusted_tis_mean": -0.4425287356321839 }, "medium": { "events": 139, "species_correct": 34, "species_attempts": 69, "withheld_or_technical": 70, "technical": 0, "species_accuracy": 0.2446043165467626, "adjusted_tis_mean": 0.8848920863309353 }, "high": { "events": 164, "species_correct": 130, "species_attempts": 152, "withheld_or_technical": 12, "technical": 0, "species_accuracy": 0.7926829268292683, "adjusted_tis_mean": 3.5121951219512195 }, "missing": { "events": 0, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 0, "technical": 0, "species_accuracy": null, "adjusted_tis_mean": null } } }, "gemini-medium": { "label": "Gemini 3.5 Flash-Lite - MEDIUM", "n_events": 390, "buckets": { "low": { "events": 43, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 43, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.06976744186046512 }, "medium": { "events": 175, "species_correct": 52, "species_attempts": 96, "withheld_or_technical": 79, "technical": 0, "species_accuracy": 0.29714285714285715, "adjusted_tis_mean": 0.9171428571428571 }, "high": { "events": 171, "species_correct": 127, "species_attempts": 142, "withheld_or_technical": 29, "technical": 0, "species_accuracy": 0.7426900584795322, "adjusted_tis_mean": 3.4210526315789473 }, "missing": { "events": 1, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 1, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-high": { "label": "Gemini 3.5 Flash-Lite - HIGH", "n_events": 390, "buckets": { "low": { "events": 36, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 36, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.125 }, "medium": { "events": 162, "species_correct": 52, "species_attempts": 101, "withheld_or_technical": 61, "technical": 0, "species_accuracy": 0.32098765432098764, "adjusted_tis_mean": 0.7067901234567902 }, "high": { "events": 190, "species_correct": 130, "species_attempts": 154, "withheld_or_technical": 36, "technical": 0, "species_accuracy": 0.6842105263157895, "adjusted_tis_mean": 3.0657894736842106 }, "missing": { "events": 2, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 2, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium37": { "label": "Gemini 3.7 Flash - MEDIUM", "n_events": 390, "buckets": { "low": { "events": 29, "species_correct": 0, "species_attempts": 2, "withheld_or_technical": 27, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.3275862068965517 }, "medium": { "events": 57, "species_correct": 16, "species_attempts": 21, "withheld_or_technical": 36, "technical": 0, "species_accuracy": 0.2807017543859649, "adjusted_tis_mean": 1.5350877192982457 }, "high": { "events": 289, "species_correct": 174, "species_attempts": 209, "withheld_or_technical": 80, "technical": 0, "species_accuracy": 0.6020761245674741, "adjusted_tis_mean": 3.0605536332179932 }, "missing": { "events": 15, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 15, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini-medium38": { "label": "Gemini 3.8 Flash - MEDIUM", "n_events": 390, "buckets": { "low": { "events": 23, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 23, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.08695652173913043 }, "medium": { "events": 89, "species_correct": 29, "species_attempts": 42, "withheld_or_technical": 47, "technical": 0, "species_accuracy": 0.3258426966292135, "adjusted_tis_mean": 1.7640449438202248 }, "high": { "events": 265, "species_correct": 155, "species_attempts": 176, "withheld_or_technical": 89, "technical": 0, "species_accuracy": 0.5849056603773585, "adjusted_tis_mean": 3.209433962264151 }, "missing": { "events": 13, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 13, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini38-low": { "label": "Gemini 3.8 Flash - LOW", "n_events": 390, "buckets": { "low": { "events": 29, "species_correct": 0, "species_attempts": 3, "withheld_or_technical": 26, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.10344827586206896 }, "medium": { "events": 85, "species_correct": 23, "species_attempts": 32, "withheld_or_technical": 53, "technical": 0, "species_accuracy": 0.27058823529411763, "adjusted_tis_mean": 1.7352941176470589 }, "high": { "events": 257, "species_correct": 163, "species_attempts": 186, "withheld_or_technical": 71, "technical": 0, "species_accuracy": 0.6342412451361867, "adjusted_tis_mean": 3.357976653696498 }, "missing": { "events": 19, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 19, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "gemini38-high": { "label": "Gemini 3.8 Flash - HIGH", "n_events": 390, "buckets": { "low": { "events": 13, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 13, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.038461538461538464 }, "medium": { "events": 110, "species_correct": 53, "species_attempts": 77, "withheld_or_technical": 33, "technical": 0, "species_accuracy": 0.4818181818181818, "adjusted_tis_mean": 1.7818181818181817 }, "high": { "events": 255, "species_correct": 173, "species_attempts": 195, "withheld_or_technical": 60, "technical": 0, "species_accuracy": 0.6784313725490196, "adjusted_tis_mean": 3.3529411764705883 }, "missing": { "events": 12, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 12, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": 0.0 } } }, "qwen-repaired": { "label": "Qwen3-VL-30B Instruct - repaired", "n_events": 390, "buckets": { "low": { "events": 65, "species_correct": 0, "species_attempts": 5, "withheld_or_technical": 60, "technical": 0, "species_accuracy": 0.0, "adjusted_tis_mean": -0.3384615384615385 }, "medium": { "events": 135, "species_correct": 20, "species_attempts": 135, "withheld_or_technical": 0, "technical": 0, "species_accuracy": 0.14814814814814814, "adjusted_tis_mean": -2.0703703703703704 }, "high": { "events": 190, "species_correct": 118, "species_attempts": 190, "withheld_or_technical": 0, "technical": 0, "species_accuracy": 0.6210526315789474, "adjusted_tis_mean": 1.7052631578947368 }, "missing": { "events": 0, "species_correct": 0, "species_attempts": 0, "withheld_or_technical": 0, "technical": 0, "species_accuracy": null, "adjusted_tis_mean": null } } } } } } } } }