{
  "evidence": {
    "runId": "1770132a-b91b-47d8-94cb-309f94811d51",
    "evaluatedAt": "2026-09-27T07:56:10.135Z",
    "packageVersion": "0.7.0",
    "recipeFingerprint": "39b2c5e0938559fc1549e339ef7191bbcf833ff8f8ec51a339aee1cb29aafe1c",
    "datasetFingerprint": "7f3fc101a33cf34ad2835b081c9f6121fce4cc1a14700fc0ded14409f149d725",
    "split": "held-out",
    "mode": "live",
    "policy": {
      "minConfidence": 0.8,
      "selectedOn": {
        "runId": "94e107aa-63fe-4dff-b134-badf2e81645d",
        "datasetFingerprint": "33c7d0024a2bc2c663c0d3c3b2336e4593defd00a2c0ec893898daa12cf77a86",
        "recipeFingerprint": "39b2c5e0938559fc1549e339ef7191bbcf833ff8f8ec51a339aee1cb29aafe1c",
        "model": "jev-1.13.0",
        "split": "development"
      }
    },
    "provenance": [
      {
        "method": "author-synthetic",
        "source": "AI-authored synthetic boundary cases for jev-recipes; labels have not had independent human review.",
        "cases": 40
      }
    ]
  },
  "recipe": "model-route",
  "model": "jev-1.13.0",
  "models": [
    "jev-1.13.0"
  ],
  "cases": 40,
  "completed": 40,
  "failed": 0,
  "correct": 40,
  "wrong": 0,
  "ready": 26,
  "review": 14,
  "accuracy": 1,
  "completedAccuracy": 1,
  "accuracyInterval95": [
    0.9124,
    1
  ],
  "readyAccuracy": 1,
  "readyAccuracyInterval95": [
    0.8713,
    1
  ],
  "reviewRate": 0.35,
  "items": 80,
  "correctItems": 80,
  "itemAccuracy": 1,
  "contestedAccuracy": 1,
  "adversarialAccuracy": null,
  "calibration": [
    {
      "confidence": "0-0.5",
      "cases": 1,
      "accuracy": 1
    },
    {
      "confidence": "0.5-0.6",
      "cases": 1,
      "accuracy": 1
    },
    {
      "confidence": "0.6-0.7",
      "cases": 1,
      "accuracy": 1
    },
    {
      "confidence": "0.7-0.8",
      "cases": 5,
      "accuracy": 1
    },
    {
      "confidence": "0.8-0.9",
      "cases": 8,
      "accuracy": 1
    },
    {
      "confidence": "0.9-1",
      "cases": 24,
      "accuracy": 1
    }
  ],
  "thresholdEvaluation": "frozen-policy",
  "thresholds": [],
  "suggestedMinConfidence": null,
  "latencyMs": {
    "p50": 124.34,
    "p95": 180.04,
    "totalCaseTime": 5181.61
  },
  "usage": {
    "requests": 40,
    "input_tokens": 30807,
    "output_tokens": 2610,
    "unknownResponses": 0,
    "failedRequests": 0,
    "models": [
      "jev-1.13.0"
    ]
  },
  "cost": {
    "model": "jev-1.13.0",
    "inputPerMillion": 0.042,
    "outputPerMillion": 0,
    "currency": "USD",
    "date": "2026-09-27",
    "source": "https://docs.typesafe.ai/models",
    "estimated": 0.00129389
  },
  "confusion": {
    "{\"suggestedSelection\":\"engineer\",\"verdict\":\"matched\"}": {
      "total": 10,
      "correct": 10,
      "misses": {}
    },
    "{\"suggestedSelection\":\"format\",\"verdict\":\"matched\"}": {
      "total": 10,
      "correct": 10,
      "misses": {}
    },
    "{\"suggestedSelection\":null,\"verdict\":\"ambiguous\"}": {
      "total": 10,
      "correct": 10,
      "misses": {}
    },
    "{\"suggestedSelection\":null,\"verdict\":\"none\"}": {
      "total": 10,
      "correct": 10,
      "misses": {}
    }
  },
  "failures": [],
  "acceptance": {
    "minConfidence": 0.8,
    "minimumHeldOutReadyCases": 20,
    "minimumReadyAccuracy": 0.95,
    "maximumProviderFailures": 0,
    "met": true,
    "label": "Measured on these synthetic cases"
  }
}