{
  "schema": "voirdire-calibration-status/1",
  "status_type": "diagnostic",
  "gate": "UNDECIDABLE",
  "gate_exit_code": 2,
  "approved_for_verdicts": false,
  "total_responses": 13230,
  "held_out_batches": 105,
  "summary": "Live calibration is complete, but it does not meet the evidence threshold for identity verdicts: Llama's 95% upper false-accusation bound is 22.38%, above the 10% limit.",
  "audited_at": "2026-10-03T13:00:29.471121+00:00",
  "collection": {
    "plan_created_at": "2026-10-03T11:21:48.420168+00:00",
    "first_response_at": "2026-10-03T11:21:52.410286+00:00",
    "last_response_at": "2026-10-03T13:00:14.402860+00:00",
    "probe_count": 21,
    "repeats_per_probe_per_model": 210,
    "temperature": 1.0,
    "max_output_tokens": 600,
    "models": {
      "gpt-class": {
        "model_id": "openai/gpt-4o-mini",
        "provider": "OpenAI",
        "endpoint_tag": "openai",
        "successful_responses": 4410,
        "finish_reasons": {
          "stop": 4410
        }
      },
      "llama-class": {
        "model_id": "meta-llama/llama-3.3-70b-instruct",
        "provider": "Groq",
        "endpoint_tag": "groq",
        "successful_responses": 4410,
        "finish_reasons": {
          "stop": 4310,
          "length": 100
        }
      },
      "mistral-class": {
        "model_id": "mistralai/mistral-small-3.2-24b-instruct",
        "provider": "Mistral",
        "endpoint_tag": "mistral/eu",
        "successful_responses": 4410,
        "finish_reasons": {
          "stop": 3815,
          "length": 594,
          "tool_calls": 1
        }
      }
    }
  },
  "evaluation": {
    "method": "disjoint-batches-heldout-v1",
    "feature_width": 112,
    "responses_per_probe_batch": 3,
    "split": "Even batch indices train; odd batch indices are held out.",
    "training_raw_responses": 6615,
    "held_out_raw_responses": 6615,
    "raw_response_overlap": 0,
    "training_batches_per_model": 35,
    "held_out_batches_per_model": 35,
    "unique_response_ids": 13230,
    "accuracy": {
      "correct": 102,
      "total": 105,
      "point_estimate": 0.9714285714285714,
      "interval_95": [
        0.9193437093626844,
        0.9902361609612977
      ]
    },
    "confusion": {
      "gpt-class": {
        "gpt-class": 35,
        "llama-class": 0,
        "mistral-class": 0
      },
      "llama-class": {
        "gpt-class": 0,
        "llama-class": 32,
        "mistral-class": 3
      },
      "mistral-class": {
        "gpt-class": 0,
        "llama-class": 0,
        "mistral-class": 35
      }
    },
    "false_accusation": {
      "gpt-class": {
        "errors": 0,
        "held_out_batches": 35,
        "point_estimate": 0.0,
        "interval_95": [
          6.938893903907228e-18,
          0.09890099232440402
        ]
      },
      "llama-class": {
        "errors": 3,
        "held_out_batches": 35,
        "point_estimate": 0.08571428571428572,
        "interval_95": [
          0.02958236826696982,
          0.22379273965896496
        ]
      },
      "mistral-class": {
        "errors": 0,
        "held_out_batches": 35,
        "point_estimate": 0.0,
        "interval_95": [
          6.938893903907228e-18,
          0.09890099232440402
        ]
      }
    },
    "error_summary": "All three held-out errors classified Llama as Mistral.",
    "confidence_method": "Two-sided 95% Wilson intervals, separately for each metric/family; not simultaneous family-wise coverage."
  },
  "thresholds": {
    "min_accuracy_lower_95": 0.75,
    "max_false_accusation_upper_95_per_family": 0.1
  },
  "gate_reasons": [
    "95% false accusation upper bound for llama-class is 0.224, above 0.100"
  ],
  "integrity": {
    "status": "COMPLETE_AND_REPRODUCIBLE",
    "complete_expected_coverage": true,
    "duplicate_upstream_response_ids": 0,
    "raw_train_test_overlap": 0,
    "matrix_reproduced_exactly": true
  },
  "sha256": {
    "corpus": "41460c34337545dc2cfb147cde5f63246a191885db1cc18c5590c7f15bd3e905",
    "plan_file": "1ec90f00ade11a19bc3153849f5ac2f4b092f17406cafcb1100629160d7e23ba",
    "plan_canonical_json": "3afc59a3d8de3273f652f2716dc8e37538295824f0d77422d72bdff60e0d0ec3",
    "diagnostic_matrix": "b6efcf6cbef0e3c01fa23d34cde353962adc2a306c8e72c0c27e3e1472e11ba2",
    "merged_response_dataset": "1562a7e25234adec12971b9b9b0d88a69d6705cc9d4253a04c39f78ff543d466",
    "per_family_response_files": {
      "gpt-class": "cfffdd8a5df0547c6401370863f9df94b17769289fe9b5a7ef4cabca6e69611b",
      "llama-class": "f3c29a68ca884a24b765b8c50aa82a20b3d541663f7f81a2737848936b4837fe",
      "mistral-class": "cfabb3a0bba85c9394e5f365cbc874dfb26585caf6f3593bb66dd51bba2e8221"
    },
    "analysis_sources": {
      "build_matrix.py": "ae851f40e769a7ec7e7b809dedfa33d039f71d2c53d92b079e2dc77ee9540626",
      "check_matrix.py": "2f6e372f68d92b5c83053c5b6f1d84d136e5fb05396bdbc53dd6a4e537c62057",
      "features.py": "f656beee4ab4b78c7453e30bd1fb591a40426c8998e3416e9d99e000b5aab008",
      "merge_runs.py": "9fe9adb14718d37fe3527a2c401044ccec76a1efae104f94b41f8284306bd1dc",
      "openrouter_battery.py": "bfc4b724d023006b2cb758ee0d12b31fd951ad80de8563f2c80d0069719989fb",
      "audit_report.py": "69b10c3b660bc640c41290859fc1c12364536a44709bb39407a3a0d38a588020"
    }
  },
  "limitations": [
    "This diagnostic report is not an approved matrix and does not authorize identity verdicts.",
    "Results cover only the listed model IDs, pinned providers, prompts, temperature and 600-token output cap.",
    "Output-cap truncations were retained: 100 Llama responses and 594 Mistral responses. One Mistral response reported tool_calls with nonempty text; it was also retained.",
    "The original plan was saved before sampling; evaluator source hashes were recorded at analysis time rather than embedded in that plan.",
    "Provider and model IDs are OpenRouter-reported metadata, not cryptographic proof of model weights.",
    "Raw response text and billing journals are not published in this status report; hashes alone do not permit independent recomputation."
  ],
  "next_step": "Preregister a new independent evaluation of a fixed abstention rule or revised corpus/classifier. Keep the current evaluated data separate from new confirmation data, and apply the unchanged evidence thresholds before approving verdicts."
}