{
 "schemaVersion": "blobfish.typed-decision-bench-runs.v1",
 "observedAt": "2026-09-19T07:08:36Z",
 "benchmark": {
  "id": "typed-decision-bench",
  "name": "Typed Decision Bench",
  "version": "0.2.0",
  "items": 6162,
  "suites": [
   "data-and-operations",
   "real-time-and-agents",
   "retrieval-and-knowledge",
   "safety-and-quality",
   "workflow-control"
  ],
  "pageHref": "/benchmarks/typed-decision-bench",
  "evaluator": "Ground-truth labels; proper scoring of returned probabilities (no LLM judge)"
 },
 "runs": [
  {
   "system": {
    "name": "OpenJev",
    "model": "openjev-latest",
    "provider": "hosted API (api.codiv.ai, free tier)",
    "repo": "razorback16/openjev",
    "revision": null,
    "variant": "zero-shot DiffusionGemma-26B server",
    "dtype": null,
    "device": "hosted",
    "temperature": null,
    "maxInputTokens": null,
    "licence": "Apache-2.0 (code)",
    "note": "Zero-shot open replica served by a third party; no task training."
   },
   "harness": "jevfish_bench runner · hosted API",
   "finishedAt": "2026-09-19T07:08:36Z",
   "aggregate": {
    "items": 6162,
    "answered": 6162,
    "skipped": 0,
    "truncated": 0,
    "correct": 4394,
    "decisionScore": 75.37,
    "accuracy": 0.7131,
    "auroc": null,
    "ece": null,
    "logLoss": null,
    "brier": null
   },
   "suites": {
    "data-and-operations": {
     "items": 1378,
     "answered": 1378,
     "skipped": 0,
     "truncated": 0,
     "correct": 1025,
     "decisionScore": 74.74,
     "accuracy": 0.7438,
     "auroc": 0.5598,
     "ece": 0.0782,
     "logLoss": 0.3403,
     "brier": 0.0653,
     "contamination": "unknown"
    },
    "real-time-and-agents": {
     "items": 1177,
     "answered": 1177,
     "skipped": 0,
     "truncated": 0,
     "correct": 681,
     "decisionScore": 68.31,
     "accuracy": 0.5786,
     "auroc": 0.4709,
     "ece": 0.2286,
     "logLoss": 2.9949,
     "brier": 0.4577,
     "contamination": "unknown"
    },
    "retrieval-and-knowledge": {
     "items": 1249,
     "answered": 1249,
     "skipped": 0,
     "truncated": 0,
     "correct": 942,
     "decisionScore": 76.45,
     "accuracy": 0.7542,
     "auroc": 0.8449,
     "ece": 0.1493,
     "logLoss": 0.9756,
     "brier": 0.1894,
     "contamination": "unknown"
    },
    "safety-and-quality": {
     "items": 1162,
     "answered": 1162,
     "skipped": 0,
     "truncated": 0,
     "correct": 959,
     "decisionScore": 85.87,
     "accuracy": 0.8253,
     "auroc": 0.9721,
     "ece": 0.1597,
     "logLoss": 0.6376,
     "brier": 0.1035,
     "contamination": "unknown"
    },
    "workflow-control": {
     "items": 1196,
     "answered": 1196,
     "skipped": 0,
     "truncated": 0,
     "correct": 787,
     "decisionScore": 71.72,
     "accuracy": 0.658,
     "auroc": 0.9681,
     "ece": 0.3628,
     "logLoss": 0.324,
     "brier": 0.0672,
     "contamination": "unknown"
    }
   },
   "provenanceHref": null
  }
 ]
}
