{
  "version": 1,
  "created_at": "2026-09-29T09:17:00.954368+00:00",
  "model": {
    "provider": "local",
    "name": "models/Qwen--Qwen3.5-0.8B",
    "adapter": "runs/decision-0.8b-v2/adapter",
    "identity_version": 2,
    "prompt_sha256": "b2c4baf03d0224bf37ea1b2e11c2fa2c98c26d09a7a881a4565150829d1e2af0",
    "model_weights_sha256": "79353a8fb811beddb052c75f92debca361226dcb0700d8f46004a77bae214f7c",
    "model_config_sha256": "779572e46ae6552ba1826a162ab198560403ab1adeed8654e24e45c14e8ba978",
    "tokenizer_sha256": "4773ccc9cb7087ffa57977144fc5786e4c1b5ef4d2b25a850bd25d3318581142",
    "snapshot_sha256": "6aa297f2cadde0920e13ed5e2e0b78a33a36046d0d5fcb8efd8b03b4cd7e5ab3",
    "weight_identity_status": "verified_local_content",
    "encoding_sha256": "1e1255124fb4f698cabd531cbc251e2b0b91b8d119a53cf88974792f196a2c30",
    "inference_dtype": "bfloat16",
    "adapter_artifacts": {
      "config_sha256": "9fee8484ccc91ee15604c0264461eecfa46b2f5605362c662d31191c08ed900e",
      "weights_sha256": "cb80a8e8a5ffd1232abaecd37a3f91a58c0f268d218069ce4cb779f70158aff8"
    }
  },
  "dataset_sha256": "7773448b87a0c4fc8e4cca7be136e1494fa9bffd5b36ae399be7e055243227d3",
  "cases_sha256": "dcc091e4687e649bd19f21e02b38a187a6b73ee909fac228e02f2da81010d631",
  "allow_test": true,
  "max_length": 2048,
  "calibration": null,
  "temperature": 1.0,
  "summary": {
    "count": 2800,
    "attempted": 2800,
    "scored": 2800,
    "errors": 0,
    "skipped": 0,
    "accuracy_all": 0.6882142857142857,
    "accuracy_attempted": 0.6882142857142857,
    "accuracy_scored": 0.6882142857142857,
    "accuracy_family_macro": 0.6261111111111111,
    "accuracy_by_family": {
      "clinc150": 0.9575,
      "exclusive_proof_fallback": 0.35,
      "go_emotions": 0.8425,
      "helpsteer2": 0.495,
      "narrow_veto_override": 0.48,
      "paws": 0.9125,
      "priority_exception": 0.33,
      "vitaminc": 0.7925,
      "weighted_role_quorum": 0.475
    },
    "by_source": {
      "clinc150": {
        "count": 400,
        "scored": 400,
        "accuracy_all": 0.9575,
        "accuracy_attempted": 0.9575,
        "macro_f1_scored": 0.9596451786927224,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "go_emotions": {
        "count": 400,
        "scored": 400,
        "accuracy_all": 0.8425,
        "accuracy_attempted": 0.8425,
        "macro_f1_scored": 0.8420646908040286,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "helpsteer2": {
        "count": 400,
        "scored": 400,
        "accuracy_all": 0.495,
        "accuracy_attempted": 0.495,
        "macro_f1_scored": 0.3669844359980635,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "paws": {
        "count": 400,
        "scored": 400,
        "accuracy_all": 0.9125,
        "accuracy_attempted": 0.9125,
        "macro_f1_scored": 0.9118249094630766,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "vitaminc": {
        "count": 400,
        "scored": 400,
        "accuracy_all": 0.7925,
        "accuracy_attempted": 0.7925,
        "macro_f1_scored": 0.7376537004098678,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "workflow.exclusive_proof_fallback": {
        "count": 200,
        "scored": 200,
        "accuracy_all": 0.35,
        "accuracy_attempted": 0.35,
        "macro_f1_scored": 0.21238337574215435,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "workflow.narrow_veto_override": {
        "count": 200,
        "scored": 200,
        "accuracy_all": 0.48,
        "accuracy_attempted": 0.48,
        "macro_f1_scored": 0.4445966514459665,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "workflow.priority_exception": {
        "count": 200,
        "scored": 200,
        "accuracy_all": 0.33,
        "accuracy_attempted": 0.33,
        "macro_f1_scored": 0.16666666666666666,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      },
      "workflow.weighted_role_quorum": {
        "count": 200,
        "scored": 200,
        "accuracy_all": 0.475,
        "accuracy_attempted": 0.475,
        "macro_f1_scored": 0.4066951566951567,
        "macro_f1_note": "Semantic option keys on successful cases; undefined for keys without shared semantics."
      }
    },
    "probability_metrics": {
      "count": 2800,
      "brier_multiclass_sum": 0.4577625055759041,
      "nll": 0.8419201488883272,
      "infinite_nll_cases": 0,
      "ece": 0.13689068341361624,
      "ece_equal_width_bins": 15,
      "aurc": 0.1427349504890547,
      "coverage_at_error": {
        "0.01": 0.14107142857142857,
        "0.05": 0.245,
        "0.1": 0.40214285714285714
      },
      "risk_coverage_note": "Empirical diagnostic, not a held-out error guarantee; whole ties accepted."
    },
    "probability_incomplete": 0,
    "latency_ms": {
      "count": 2800,
      "mean": 19.245112702062347,
      "p50": 19.073196999670472,
      "p95": 19.319465538865188,
      "min": 18.60275999933947,
      "max": 415.4445160093019
    },
    "wall_latency_ms": {
      "count": 2800,
      "mean": 21.053149590614858,
      "p50": 20.84802999161184,
      "p95": 23.9287625576253,
      "min": 19.18114699947182,
      "max": 437.42106699210126
    },
    "input_tokens": {
      "count": 2800,
      "mean": 334.02714285714285,
      "p50": 276.0,
      "p95": 844.0,
      "min": 108.0,
      "max": 1746.0
    },
    "state_characters": {
      "count": 2800,
      "mean": 482.25107142857144,
      "p50": 206.0,
      "p95": 2657.899999999995,
      "min": 7.0,
      "max": 6978.0
    },
    "notes": [
      "accuracy_all counts errors and skips as incorrect",
      "accuracy_attempted and family macro count errors as wrong and exclude intentional skips",
      "family defaults to source; latency_ms is local forward-pass or hosted HTTP time",
      "wall_latency_ms includes per-record scorer preprocessing and inference; no warmup excluded",
      "F1 uses semantic option keys, never shuffled answer letters",
      "Probability metrics only include explicitly complete distributions"
    ]
  },
  "split": "test",
  "source_report_sha256": "4ccfb86a7c6df69cf94efb8aa6aa45d16f373cd41a3916879b9db12b53bdb85c",
  "source_report_name": "historical-decision.json",
  "scope": "Frozen candidate test report; independence depends on recorded protocol"
}