{
  "runId": "c99506be-e7cc-4900-90b0-2a059597aa89",
  "generatedAt": "2026-07-10T16:21:33.096Z",
  "models": [
    {
      "adapterId": "agy",
      "model": "Gemini 3.5 Flash (High)",
      "label": "agy/Gemini 3.5 Flash (High)",
      "attempts": 38,
      "scored": 37,
      "pass": 32,
      "partial": 3,
      "incorrect": 2,
      "protocol_error": 1,
      "otherErrors": 0,
      "passRate": 0.8648648648648649,
      "quality": 0.9054054054054054,
      "coverage": 0.9736842105263158,
      "fixtureScore": 0.881578947368421,
      "fixtureSolveRate": 0.8421052631578947,
      "passAtKRate": 0.8648648648648649,
      "fixturesPassed": 32,
      "fixturesAttempted": 38,
      "avgDurationMs": 6920,
      "minDurationMs": 4475,
      "maxDurationMs": 37526,
      "totalDurationMs": 262959,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 1,
      "passRateLabel": "86.5%",
      "qualityLabel": "0.905",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "88.2%",
      "fixtureSolveRateLabel": "84.2%",
      "passAtKLabel": "86.5%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.1 Pro (Low)",
      "label": "agy/Gemini 3.1 Pro (Low)",
      "attempts": 38,
      "scored": 35,
      "pass": 31,
      "partial": 4,
      "incorrect": 0,
      "protocol_error": 3,
      "otherErrors": 0,
      "passRate": 0.8857142857142857,
      "quality": 0.9428571428571428,
      "coverage": 0.9210526315789473,
      "fixtureScore": 0.868421052631579,
      "fixtureSolveRate": 0.8157894736842105,
      "passAtKRate": 0.8857142857142857,
      "fixturesPassed": 31,
      "fixturesAttempted": 38,
      "avgDurationMs": 11095,
      "minDurationMs": 2346,
      "maxDurationMs": 16802,
      "totalDurationMs": 421591,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 2,
      "passRateLabel": "88.6%",
      "qualityLabel": "0.943",
      "coverageLabel": "92.1%",
      "fixtureScoreLabel": "86.8%",
      "fixtureSolveRateLabel": "81.6%",
      "passAtKLabel": "88.6%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.5 Flash (Low)",
      "label": "agy/Gemini 3.5 Flash (Low)",
      "attempts": 38,
      "scored": 37,
      "pass": 29,
      "partial": 5,
      "incorrect": 3,
      "protocol_error": 1,
      "otherErrors": 0,
      "passRate": 0.7837837837837838,
      "quality": 0.8513513513513513,
      "coverage": 0.9736842105263158,
      "fixtureScore": 0.8289473684210527,
      "fixtureSolveRate": 0.7631578947368421,
      "passAtKRate": 0.7837837837837838,
      "fixturesPassed": 29,
      "fixturesAttempted": 38,
      "avgDurationMs": 6355,
      "minDurationMs": 3756,
      "maxDurationMs": 67432,
      "totalDurationMs": 241498,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 3,
      "passRateLabel": "78.4%",
      "qualityLabel": "0.851",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "82.9%",
      "fixtureSolveRateLabel": "76.3%",
      "passAtKLabel": "78.4%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.5 Flash (Medium)",
      "label": "agy/Gemini 3.5 Flash (Medium)",
      "attempts": 38,
      "scored": 37,
      "pass": 29,
      "partial": 5,
      "incorrect": 3,
      "protocol_error": 1,
      "otherErrors": 0,
      "passRate": 0.7837837837837838,
      "quality": 0.8513513513513513,
      "coverage": 0.9736842105263158,
      "fixtureScore": 0.8289473684210527,
      "fixtureSolveRate": 0.7631578947368421,
      "passAtKRate": 0.7837837837837838,
      "fixturesPassed": 29,
      "fixturesAttempted": 38,
      "avgDurationMs": 5903,
      "minDurationMs": 4289,
      "maxDurationMs": 9313,
      "totalDurationMs": 224301,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 4,
      "passRateLabel": "78.4%",
      "qualityLabel": "0.851",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "82.9%",
      "fixtureSolveRateLabel": "76.3%",
      "passAtKLabel": "78.4%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.1 Pro (High)",
      "label": "agy/Gemini 3.1 Pro (High)",
      "attempts": 38,
      "scored": 27,
      "pass": 22,
      "partial": 3,
      "incorrect": 2,
      "protocol_error": 11,
      "otherErrors": 0,
      "passRate": 0.8148148148148148,
      "quality": 0.8703703703703703,
      "coverage": 0.7105263157894737,
      "fixtureScore": 0.618421052631579,
      "fixtureSolveRate": 0.5789473684210527,
      "passAtKRate": 0.8148148148148148,
      "fixturesPassed": 22,
      "fixturesAttempted": 38,
      "avgDurationMs": 11498,
      "minDurationMs": 5992,
      "maxDurationMs": 25988,
      "totalDurationMs": 436909,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 5,
      "passRateLabel": "81.5%",
      "qualityLabel": "0.870",
      "coverageLabel": "71.1%",
      "fixtureScoreLabel": "61.8%",
      "fixtureSolveRateLabel": "57.9%",
      "passAtKLabel": "81.5%"
    },
    {
      "adapterId": "agy",
      "model": "Claude Opus 4.6 (Thinking)",
      "label": "agy/Claude Opus 4.6 (Thinking)",
      "attempts": 38,
      "scored": 0,
      "pass": 0,
      "partial": 0,
      "incorrect": 0,
      "protocol_error": 38,
      "otherErrors": 0,
      "passRate": 0,
      "quality": 0,
      "coverage": 0,
      "fixtureScore": 0,
      "fixtureSolveRate": 0,
      "passAtKRate": 0,
      "fixturesPassed": 0,
      "fixturesAttempted": 38,
      "avgDurationMs": 6798,
      "minDurationMs": 5548,
      "maxDurationMs": 10236,
      "totalDurationMs": 258335,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 6,
      "passRateLabel": "0.0%",
      "qualityLabel": "0.000",
      "coverageLabel": "0.0%",
      "fixtureScoreLabel": "0.0%",
      "fixtureSolveRateLabel": "0.0%",
      "passAtKLabel": "0.0%"
    },
    {
      "adapterId": "agy",
      "model": "Claude Sonnet 4.6 (Thinking)",
      "label": "agy/Claude Sonnet 4.6 (Thinking)",
      "attempts": 38,
      "scored": 0,
      "pass": 0,
      "partial": 0,
      "incorrect": 0,
      "protocol_error": 38,
      "otherErrors": 0,
      "passRate": 0,
      "quality": 0,
      "coverage": 0,
      "fixtureScore": 0,
      "fixtureSolveRate": 0,
      "passAtKRate": 0,
      "fixturesPassed": 0,
      "fixturesAttempted": 38,
      "avgDurationMs": 6709,
      "minDurationMs": 5734,
      "maxDurationMs": 11740,
      "totalDurationMs": 254945,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 7,
      "passRateLabel": "0.0%",
      "qualityLabel": "0.000",
      "coverageLabel": "0.0%",
      "fixtureScoreLabel": "0.0%",
      "fixtureSolveRateLabel": "0.0%",
      "passAtKLabel": "0.0%"
    },
    {
      "adapterId": "agy",
      "model": "GPT-OSS 120B (Medium)",
      "label": "agy/GPT-OSS 120B (Medium)",
      "attempts": 38,
      "scored": 0,
      "pass": 0,
      "partial": 0,
      "incorrect": 0,
      "protocol_error": 38,
      "otherErrors": 0,
      "passRate": 0,
      "quality": 0,
      "coverage": 0,
      "fixtureScore": 0,
      "fixtureSolveRate": 0,
      "passAtKRate": 0,
      "fixturesPassed": 0,
      "fixturesAttempted": 38,
      "avgDurationMs": 7179,
      "minDurationMs": 5987,
      "maxDurationMs": 11099,
      "totalDurationMs": 272788,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 8,
      "passRateLabel": "0.0%",
      "qualityLabel": "0.000",
      "coverageLabel": "0.0%",
      "fixtureScoreLabel": "0.0%",
      "fixtureSolveRateLabel": "0.0%",
      "passAtKLabel": "0.0%"
    }
  ]
}