{
  "evidence": {
    "runId": "4b0b0786-10ac-4f1f-8399-e7a67b0cd9e1",
    "evaluatedAt": "2026-09-27T07:56:15.276Z",
    "packageVersion": "0.7.0",
    "recipeFingerprint": "4325f3b69b7412e092d660cb8a88b35faf53642a1310dff64aabd0557ca88f46",
    "datasetFingerprint": "d561d098f8fa6b63ebe72dc1ad83acd4958643f9b19e4c7dd64c6ae611dcc236",
    "split": "held-out",
    "mode": "live",
    "policy": {
      "minConfidence": 0.8,
      "selectedOn": {
        "runId": "b6439654-cf9e-4cf5-8838-ac91e67a478e",
        "datasetFingerprint": "999616ae8c86914c756a3ba0478d89b4e999bd236eff98c8546017fba046e86b",
        "recipeFingerprint": "4325f3b69b7412e092d660cb8a88b35faf53642a1310dff64aabd0557ca88f46",
        "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": "retry-worthwhile",
  "model": "jev-1.13.0",
  "models": [
    "jev-1.13.0"
  ],
  "cases": 40,
  "completed": 40,
  "failed": 0,
  "correct": 40,
  "wrong": 0,
  "ready": 39,
  "review": 1,
  "accuracy": 1,
  "completedAccuracy": 1,
  "accuracyInterval95": [
    0.9124,
    1
  ],
  "readyAccuracy": 1,
  "readyAccuracyInterval95": [
    0.9103,
    1
  ],
  "reviewRate": 0.025,
  "contestedAccuracy": null,
  "adversarialAccuracy": null,
  "calibration": [
    {
      "confidence": "0-0.5",
      "cases": 0,
      "accuracy": null
    },
    {
      "confidence": "0.5-0.6",
      "cases": 0,
      "accuracy": null
    },
    {
      "confidence": "0.6-0.7",
      "cases": 0,
      "accuracy": null
    },
    {
      "confidence": "0.7-0.8",
      "cases": 1,
      "accuracy": 1
    },
    {
      "confidence": "0.8-0.9",
      "cases": 3,
      "accuracy": 1
    },
    {
      "confidence": "0.9-1",
      "cases": 36,
      "accuracy": 1
    }
  ],
  "thresholdEvaluation": "frozen-policy",
  "thresholds": [],
  "suggestedMinConfidence": null,
  "latencyMs": {
    "p50": 109,
    "p95": 136.32,
    "totalCaseTime": 4412.87
  },
  "usage": {
    "requests": 40,
    "input_tokens": 18828,
    "output_tokens": 800,
    "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": 0.00079078
  },
  "confusion": {
    "{\"verdict\":\"retry\"}": {
      "total": 20,
      "correct": 20,
      "misses": {}
    },
    "{\"verdict\":\"stop\"}": {
      "total": 20,
      "correct": 20,
      "misses": {}
    }
  },
  "failures": [],
  "acceptance": {
    "minConfidence": 0.8,
    "minimumHeldOutReadyCases": 20,
    "minimumReadyAccuracy": 0.95,
    "maximumProviderFailures": 0,
    "met": true,
    "label": "Measured on these synthetic cases"
  }
}