{
  "runId": "c4a59a56-eb25-4438-afae-b13af70d3626",
  "generatedAt": "2026-07-11T06:53:24.362Z",
  "models": [
    {
      "adapterId": "agy",
      "model": "Claude Sonnet 4.6 (Thinking)",
      "label": "agy/Claude Sonnet 4.6 (Thinking)",
      "attempts": 38,
      "scored": 35,
      "pass": 29,
      "partial": 2,
      "incorrect": 4,
      "protocol_error": 3,
      "otherErrors": 0,
      "passRate": 0.8285714285714286,
      "quality": 0.8571428571428571,
      "coverage": 0.9210526315789473,
      "fixtureScore": 0.7894736842105263,
      "fixtureSolveRate": 0.7631578947368421,
      "passAtKRate": 0.8285714285714286,
      "fixturesPassed": 29,
      "fixturesAttempted": 38,
      "avgDurationMs": 9886,
      "minDurationMs": 6516,
      "maxDurationMs": 18770,
      "totalDurationMs": 375649,
      "totalPromptTokens": 2761,
      "totalCompletionTokens": 5312,
      "totalReasoningTokens": 0,
      "totalTokens": 8073,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 21258,
      "rank": 1,
      "passRateLabel": "82.9%",
      "qualityLabel": "0.857",
      "coverageLabel": "92.1%",
      "fixtureScoreLabel": "78.9%",
      "fixtureSolveRateLabel": "76.3%",
      "passAtKLabel": "82.9%"
    },
    {
      "adapterId": "agy",
      "model": "Claude Opus 4.6 (Thinking)",
      "label": "agy/Claude Opus 4.6 (Thinking)",
      "attempts": 38,
      "scored": 12,
      "pass": 10,
      "partial": 1,
      "incorrect": 1,
      "protocol_error": 26,
      "otherErrors": 0,
      "passRate": 0.8333333333333334,
      "quality": 0.875,
      "coverage": 0.3157894736842105,
      "fixtureScore": 0.27631578947368424,
      "fixtureSolveRate": 0.2631578947368421,
      "passAtKRate": 0.8333333333333334,
      "fixturesPassed": 10,
      "fixturesAttempted": 38,
      "avgDurationMs": 7306,
      "minDurationMs": 5551,
      "maxDurationMs": 12722,
      "totalDurationMs": 277614,
      "totalPromptTokens": 2761,
      "totalCompletionTokens": 1173,
      "totalReasoningTokens": 0,
      "totalTokens": 3934,
      "totalCachedTokens": 0,
      "totalCost": 0,
      "totalAnswerBytes": 4672,
      "rank": 2,
      "passRateLabel": "83.3%",
      "qualityLabel": "0.875",
      "coverageLabel": "31.6%",
      "fixtureScoreLabel": "27.6%",
      "fixtureSolveRateLabel": "26.3%",
      "passAtKLabel": "83.3%"
    }
  ]
}