{
  "schema_version": 1,
  "at_utc": "2026-10-23T10:34:49.444468+00:00",
  "scope": "Parent reconciliation of completed anonymous reviews and frozen saved evidence. No new candidate execution, model request, retry, rescue or service action.",
  "unit": "Three trajectories in one retained application domain; four dependent user turns per trajectory. No blended test-count ranking.",
  "blind_packet": {
    "sha256": "7bc926d0dc53689258843c8bf4b18111cf14907995e307de8923547a60f5b8f2",
    "inventory_entries": 407,
    "independently_recomputed_after_both_reviews": true,
    "unchanged": true
  },
  "review_sha256": {
    "blind-code-review.json": "f620e9fc7e70b58fe1d4e753090e218608a282496fe118283f2e19aed62082fa",
    "blind-operations-review.json": "adea7628c5e11eaced3979ee9a2cbe304ef19ac22e8ec400a45b9b5b18492ee2"
  },
  "identity_join_after_both_reviews": {
    "Amber": "ornith-ultra",
    "Cobalt": "raon-1-mini-q8",
    "Juniper": "flint"
  },
  "findings": [
    {
      "id": "R-credits",
      "decision": "material defect",
      "basis": "Both reviews plus saved additive diagnostic",
      "detail": "Explicit identical manual adjustments are suppressed; saved result Bob2100/one audit row versus required1900/two rows. Persistent from turn1; count once."
    },
    {
      "id": "R-correction-recovery",
      "decision": "material composition defect",
      "basis": "Both reviews, final source and frozen migrated-end-to-end-recovery failure",
      "detail": "recover tests order-ID existence rather than application/revision/status. Frozen migrated_workflow stages newer revisions of existing L1/L2, recovers and publishes; older rows remain. Earlier isolated-mechanism passes do not cover this composition.",
      "references": [
        "blind-review-packet/Cobalt/turn-4/work/exporter.py:285",
        "fixtures/export/coding/_private/checks.py:255"
      ]
    },
    {
      "id": "R-backup-verifier",
      "decision": "material generalization defect",
      "basis": "Both independent source reviews",
      "detail": "Hardcoded sample backup rows replace actual source equivalence; src is unused and unknown objects/data are unchecked. A canonical sample backup beside a changed source can pass; legitimate different matching data can fail. Static deductions; this source-counterexample was not newly executed.",
      "qualification": "Saved invalid/altered-backup/matching-backup diagnostics all pass on their particular fixtures. This defect does not cause the frozen backup/history failure; that checker uses the same canonical fixture."
    },
    {
      "id": "R-initial-history-pointer",
      "decision": "real bookkeeping gap; reduced weight due less explicit immediate pointer expectation",
      "basis": "Recorded empty latest immediately after migration, source and frozen migration-backup-and-history checker",
      "detail": "Backup version/count/note assertions succeed. assert_report defaults latest=True, and migration registers published history without latest/pointer. Later explicit publish yields correct retained final state. No lost backup or historical bytes demonstrated.",
      "references": [
        "fixtures/export/coding/_private/checks.py:230",
        "fixtures/export/coding/_private/oracle.py"
      ]
    },
    {
      "id": "R-transaction-boundary",
      "decision": "source-proven atomicity boundary; no observed resulting data loss",
      "basis": "Code-review C5 and source",
      "detail": "with db does not explicitly begin before first CREATE meta; following INSERT begins implicit DML transaction. Later failures can leave first DDL outside rollback. Not fault-injected; do not describe as observed ledger erasure."
    },
    {
      "id": "R-self-correction",
      "decision": "retain positive evidence and historical breach separately",
      "basis": "Recorded operations and both reviews",
      "detail": "Intermediate cascade deleted credit and published4600. Model inspected audit, fixed upsert, restored credit and original id1, delivered4400 and later disclosed restoration. Erroneous historical edition remains; no permanent id loss claimed."
    },
    {
      "id": "F-source-evidence",
      "decision": "material retained-operation defect; count once",
      "basis": "Both reviews, public/source differences and split retained diagnostic",
      "detail": "New Alice revision3 changed to2; old1400/rev2 changed to1200/rev1. Same persistent source violation across turns2/3, with other retained assertions and fresh mechanisms passing."
    },
    {
      "id": "F-manual-reset",
      "decision": "out-of-contract intervention, qualified",
      "basis": "Code-review correction and immediate preimage tool24",
      "detail": "Tool25 broad DELETE/reset/reopen violates no-reset instruction. Affected tables already empty and generation0; staged source rows survived. No erased nonempty rows demonstrated. Do not promote this to observed data destruction."
    },
    {
      "id": "F-partial-regressions",
      "decision": "observed unfinished-work regression; no completed migration accuracy score",
      "basis": "Both reviews, terminal source and actual failing unittest output",
      "detail": "Turn4 restored old exporter, losing revision/upsert/audit/history repairs. Earlier turn3 implementation and substantive incident assertions remain positive evidence; loop stop precludes migration verdict."
    },
    {
      "id": "loops",
      "decision": "two confirmed loops ended; unavailable turns unscored",
      "basis": "Operational loop audits and canonical terminal receipts",
      "detail": "Ornith32 repeats settled policy oscillations despite feasible SQLite probes; Flint67 repeats stale chronology after conclusive source inspection. No arbitrary timeout explanation; no causal attribution isolating weights, quantization or runtime."
    }
  ],
  "recommendation": {
    "order_for_sustained_work": [
      "raon-1-mini-q8",
      "ornith-ultra",
      "flint"
    ],
    "confidence": "moderate for practical long-task preference, low for broad superiority",
    "reason": "Raon alone demonstrated continued useful action through the complete retained sequence and real handoff, while preserving supplied sources. Its three material silent defects prevent a clean correctness win. Earlier six-form panel favors Ornith short-repair correctness and gives counterevidence against broad dismissal of all three. Flint speed and completed incident skill do not erase evidence-integrity and regression problems.",
    "deployment": "Recommendation only; original Ornith C12 default restored. Raon/Flint native packages were evaluated, no new Halo runtime port."
  },
  "raw_checker_results_unchanged": true,
  "incomplete_or_unattempted_accuracy_zeros": false,
  "final_protocol_audit_sha256": "fe1cbfabf48564aee8294b8db74f5fc6738a878db1f4b7af4972aa47f4132229",
  "host_receipt_sha256": "9101b71a115c48eac4341fba3a759b6ec9f3303488ed21ce82570a3b314e67da",
  "further_model_calls": 0
}
