{
  "runId": "0f05b466-ff5a-469a-9def-8077a159a510",
  "generatedAt": "2026-07-10T15:53:30.573Z",
  "models": [
    {
      "adapterId": "agy",
      "model": "Gemini 3.1 Pro (High)",
      "label": "agy/Gemini 3.1 Pro (High)",
      "attempts": 38,
      "scored": 37,
      "pass": 33,
      "partial": 4,
      "incorrect": 0,
      "protocol_error": 1,
      "otherErrors": 0,
      "passRate": 0.8918918918918919,
      "quality": 0.9459459459459459,
      "coverage": 0.9736842105263158,
      "fixtureScore": 0.9210526315789473,
      "fixtureSolveRate": 0.868421052631579,
      "passAtKRate": 0.8918918918918919,
      "fixturesPassed": 33,
      "fixturesAttempted": 38,
      "avgDurationMs": 14537,
      "minDurationMs": 8254,
      "maxDurationMs": 29261,
      "totalDurationMs": 552391,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 1,
      "passRateLabel": "89.2%",
      "qualityLabel": "0.946",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "92.1%",
      "fixtureSolveRateLabel": "86.8%",
      "passAtKLabel": "89.2%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.5 Flash (Medium)",
      "label": "agy/Gemini 3.5 Flash (Medium)",
      "attempts": 38,
      "scored": 37,
      "pass": 32,
      "partial": 4,
      "incorrect": 1,
      "protocol_error": 1,
      "otherErrors": 0,
      "passRate": 0.8648648648648649,
      "quality": 0.918918918918919,
      "coverage": 0.9736842105263158,
      "fixtureScore": 0.8947368421052632,
      "fixtureSolveRate": 0.8421052631578947,
      "passAtKRate": 0.8648648648648649,
      "fixturesPassed": 32,
      "fixturesAttempted": 38,
      "avgDurationMs": 6053,
      "minDurationMs": 4645,
      "maxDurationMs": 12348,
      "totalDurationMs": 230000,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 2,
      "passRateLabel": "86.5%",
      "qualityLabel": "0.919",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "89.5%",
      "fixtureSolveRateLabel": "84.2%",
      "passAtKLabel": "86.5%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.1 Pro (Low)",
      "label": "agy/Gemini 3.1 Pro (Low)",
      "attempts": 38,
      "scored": 36,
      "pass": 32,
      "partial": 2,
      "incorrect": 2,
      "protocol_error": 2,
      "otherErrors": 0,
      "passRate": 0.8888888888888888,
      "quality": 0.9166666666666666,
      "coverage": 0.9473684210526315,
      "fixtureScore": 0.868421052631579,
      "fixtureSolveRate": 0.8421052631578947,
      "passAtKRate": 0.8888888888888888,
      "fixturesPassed": 32,
      "fixturesAttempted": 38,
      "avgDurationMs": 11456,
      "minDurationMs": 7685,
      "maxDurationMs": 19316,
      "totalDurationMs": 435334,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 3,
      "passRateLabel": "88.9%",
      "qualityLabel": "0.917",
      "coverageLabel": "94.7%",
      "fixtureScoreLabel": "86.8%",
      "fixtureSolveRateLabel": "84.2%",
      "passAtKLabel": "88.9%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.5 Flash (High)",
      "label": "agy/Gemini 3.5 Flash (High)",
      "attempts": 38,
      "scored": 37,
      "pass": 29,
      "partial": 6,
      "incorrect": 2,
      "protocol_error": 1,
      "otherErrors": 0,
      "passRate": 0.7837837837837838,
      "quality": 0.8648648648648649,
      "coverage": 0.9736842105263158,
      "fixtureScore": 0.8421052631578947,
      "fixtureSolveRate": 0.7631578947368421,
      "passAtKRate": 0.7837837837837838,
      "fixturesPassed": 29,
      "fixturesAttempted": 38,
      "avgDurationMs": 6856,
      "minDurationMs": 4456,
      "maxDurationMs": 33576,
      "totalDurationMs": 260524,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 4,
      "passRateLabel": "78.4%",
      "qualityLabel": "0.865",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "84.2%",
      "fixtureSolveRateLabel": "76.3%",
      "passAtKLabel": "78.4%"
    },
    {
      "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": 4696,
      "minDurationMs": 3730,
      "maxDurationMs": 7343,
      "totalDurationMs": 178459,
      "totalPromptTokens": 0,
      "totalCompletionTokens": 0,
      "totalReasoningTokens": 0,
      "totalTokens": 0,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 0,
      "rank": 5,
      "passRateLabel": "78.4%",
      "qualityLabel": "0.851",
      "coverageLabel": "97.4%",
      "fixtureScoreLabel": "82.9%",
      "fixtureSolveRateLabel": "76.3%",
      "passAtKLabel": "78.4%"
    },
    {
      "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": 6695,
      "minDurationMs": 6051,
      "maxDurationMs": 7350,
      "totalDurationMs": 254401,
      "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": 6543,
      "minDurationMs": 5830,
      "maxDurationMs": 7115,
      "totalDurationMs": 248617,
      "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": 6487,
      "minDurationMs": 5707,
      "maxDurationMs": 7250,
      "totalDurationMs": 246501,
      "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%"
    }
  ]
}