{
  "runId": "49efc61a-73c8-4c9c-bcf1-1efaf96c4aef",
  "generatedAt": "2026-07-10T22:52:02.956Z",
  "models": [
    {
      "adapterId": "agy",
      "model": "Gemini 3.1 Pro (High)",
      "label": "agy/Gemini 3.1 Pro (High)",
      "attempts": 285,
      "scored": 276,
      "pass": 200,
      "partial": 50,
      "incorrect": 26,
      "protocol_error": 9,
      "otherErrors": 0,
      "passRate": 0.7246376811594203,
      "quality": 0.8152173913043478,
      "coverage": 0.968421052631579,
      "fixtureScore": 0.7894736842105264,
      "fixtureSolveRate": 0.7719298245614035,
      "passAtKRate": 0.7857142857142857,
      "fixturesPassed": 44,
      "fixturesAttempted": 57,
      "avgDurationMs": 18397,
      "minDurationMs": 7147,
      "maxDurationMs": 106104,
      "totalDurationMs": 5243091,
      "totalPromptTokens": 23505,
      "totalCompletionTokens": 34476,
      "totalReasoningTokens": 0,
      "totalTokens": 57981,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 137497,
      "rank": 1,
      "passRateLabel": "72.5%",
      "qualityLabel": "0.815",
      "coverageLabel": "96.8%",
      "fixtureScoreLabel": "78.9%",
      "fixtureSolveRateLabel": "77.2%",
      "passAtKLabel": "78.6%"
    },
    {
      "adapterId": "agy",
      "model": "Gemini 3.1 Pro (Low)",
      "label": "agy/Gemini 3.1 Pro (Low)",
      "attempts": 285,
      "scored": 275,
      "pass": 191,
      "partial": 49,
      "incorrect": 35,
      "protocol_error": 10,
      "otherErrors": 0,
      "passRate": 0.6945454545454546,
      "quality": 0.7836363636363637,
      "coverage": 0.9649122807017544,
      "fixtureScore": 0.756140350877193,
      "fixtureSolveRate": 0.7368421052631579,
      "passAtKRate": 0.75,
      "fixturesPassed": 42,
      "fixturesAttempted": 57,
      "avgDurationMs": 13029,
      "minDurationMs": 7521,
      "maxDurationMs": 34188,
      "totalDurationMs": 3713321,
      "totalPromptTokens": 23505,
      "totalCompletionTokens": 34171,
      "totalReasoningTokens": 0,
      "totalTokens": 57676,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 136278,
      "rank": 2,
      "passRateLabel": "69.5%",
      "qualityLabel": "0.784",
      "coverageLabel": "96.5%",
      "fixtureScoreLabel": "75.6%",
      "fixtureSolveRateLabel": "73.7%",
      "passAtKLabel": "75.0%"
    }
  ]
}