{
  "evidence": {
    "runId": "00a3ca39-e440-4c59-82af-28f6ca9d1430",
    "evaluatedAt": "2026-09-27T07:56:14.022Z",
    "packageVersion": "0.7.0",
    "recipeFingerprint": "cd8c13f10bddc73e3eee93efc0f14fe54eef3d28c6e651e224e6b13a449df296",
    "datasetFingerprint": "f8043bd73140bfde2a098f4093f0f612feb3612df1308a442d5aab1b0ee4cc25",
    "split": "held-out",
    "mode": "live",
    "policy": {
      "minConfidence": 0.8,
      "selectedOn": {
        "runId": "742d5cec-b756-4858-8b4d-e215e487807d",
        "datasetFingerprint": "35981792515eea8c39d6038846b7f44a71a9e9f46d970fe48af261ad3d075785",
        "recipeFingerprint": "cd8c13f10bddc73e3eee93efc0f14fe54eef3d28c6e651e224e6b13a449df296",
        "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": "progress-stall",
  "model": "jev-1.13.0",
  "models": [
    "jev-1.13.0"
  ],
  "cases": 40,
  "completed": 40,
  "failed": 0,
  "correct": 40,
  "wrong": 0,
  "ready": 40,
  "review": 0,
  "accuracy": 1,
  "completedAccuracy": 1,
  "accuracyInterval95": [
    0.9124,
    1
  ],
  "readyAccuracy": 1,
  "readyAccuracyInterval95": [
    0.9124,
    1
  ],
  "reviewRate": 0,
  "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": 0,
      "accuracy": null
    },
    {
      "confidence": "0.8-0.9",
      "cases": 1,
      "accuracy": 1
    },
    {
      "confidence": "0.9-1",
      "cases": 39,
      "accuracy": 1
    }
  ],
  "thresholdEvaluation": "frozen-policy",
  "thresholds": [],
  "suggestedMinConfidence": null,
  "latencyMs": {
    "p50": 118.82,
    "p95": 159.07,
    "totalCaseTime": 4850.19
  },
  "usage": {
    "requests": 40,
    "input_tokens": 20504,
    "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": 0.00086117
  },
  "confusion": {
    "{\"verdict\":\"progressing\"}": {
      "total": 20,
      "correct": 20,
      "misses": {}
    },
    "{\"verdict\":\"stalled\"}": {
      "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"
  }
}