{
  "schema_version": 1,
  "review_type": "Anonymous static source and recorded-operation review of one sustained four-turn Export trajectory per candidate",
  "packet_inventory_sha256_supplied": "7bc926d0dc53689258843c8bf4b18111cf14907995e307de8923547a60f5b8f2",
  "packet_hash_note": "Supplied inventory identity carried into this report; the inventory hash was not independently recomputed.",
  "scope_limits": [
    "Read public prompts, interface, initial application and later evidence; anonymous natural stage snapshots, terminal partial source, recorded nonreasoning operations, and saved independent diagnostics only.",
    "No ancestor reports, identity map, model identities, timing comparisons, memory, or prior reviews were consulted.",
    "No candidate Python or SQLite was executed on the host. No new tests, reruns, rescues, model requests, Docker, service, remote actions, or packet edits were performed.",
    "Installed requesting-code-review, its code-reviewer template, and verification-before-completion skills informed requirement tracing and evidence calibration. This is the dispatched reviewer; no further reviewers were spawned.",
    "Positive runtime statements refer to saved recorded results, not fresh executions. Source counterexamples are explicitly identified as static deductions.",
    "Dependent stages are not independent samples. No blended score, pass-percentage ranking, test-count ranking, or accuracy score for incomplete or unattempted turns is assigned."
  ],
  "reference_convention": "All evidence paths below are relative to packet_root. Line numbers refer to the saved text files; operations.json references include the tool sequence for retrieval.",
  "evidence_table": [
    {
      "candidate": "Amber",
      "turn": 1,
      "availability": "completed natural snapshot",
      "source_result": "Retry applies accepted staging; recovery applies accepted batches and places/acknowledges stored prepared bodies. Explicit manual adjustments remain additive.",
      "operation_result": "Recorded retained demo has three orders, one packing-credit adjustment, and the original 3300-cent report with matching pointer.",
      "limits": "The initial replacement-based importer and one-report-per-day policy remain; later requirements were not yet completed.",
      "references": [
        "Amber/turn-1/work/exporter.py:61-89",
        "Amber/turn-1/work/exporter.py:167-173",
        "Amber/operations.json:320-388 (tool-000024 through tool-000028)"
      ]
    },
    {
      "candidate": "Amber",
      "turn": 2,
      "availability": "attempted, incomplete",
      "source_result": "Terminal exporter is unchanged from turn 1: INSERT OR REPLACE still bypasses revision precedence and can cascade-delete credits; report preparation still reuses the day key/path for changed bodies.",
      "operation_result": "Recorded exploration examines replacement versus upsert behavior, but there is no completed correction/edition implementation or operated 4400 report.",
      "limits": "These are unfinished frontier requirements, not a fabricated completed-turn verdict.",
      "references": [
        "Amber/terminal-partial-work/exporter.py:51-58",
        "Amber/terminal-partial-work/exporter.py:115-131",
        "Amber/operations.json:502-540 (tool-000037 through tool-000039)"
      ]
    },
    {
      "candidate": "Amber",
      "turn": 3,
      "availability": "unattempted",
      "source_result": null,
      "operation_result": null,
      "limits": "No overlap-repair verdict or accuracy score is available.",
      "references": [
        "manifest.json:8-24"
      ]
    },
    {
      "candidate": "Amber",
      "turn": 4,
      "availability": "unattempted",
      "source_result": null,
      "operation_result": null,
      "limits": "No migration or handoff verdict or accuracy score is available.",
      "references": [
        "manifest.json:8-24"
      ]
    },
    {
      "candidate": "Cobalt",
      "turn": 1,
      "availability": "completed natural snapshot",
      "source_result": "Initial accepted-order recovery and retry are repaired, including rescue of staged rows whose import marker was already falsely marked done. Explicit identical adjustments are incorrectly deduplicated, and retry returns its stale pre-application accepted status.",
      "operation_result": "The retained initial ledger/report reaches 3300. Saved stage checks support the narrow initial recovery case, not the changed additive-adjustment contract.",
      "limits": "The deduplication defect persists through the completed fourth turn.",
      "references": [
        "Cobalt/turn-1/work/exporter.py:61-97",
        "Cobalt/turn-1/work/exporter.py:175-194",
        "Cobalt/operations.json:284-338 (tool-000021 through tool-000024)",
        "Cobalt/turn-4/public-contract-diagnostics.json:3-44"
      ]
    },
    {
      "candidate": "Cobalt",
      "turn": 2,
      "availability": "completed natural snapshot",
      "source_result": "Final source updates orders in place only for greater revisions and makes digest-scoped editions while reusing identical historical content. Recovery still tests only order-ID presence, so accepted corrections to existing orders are skipped.",
      "operation_result": "4400 and original 3300 reports are retained. During operation an intermediate implementation deleted the packing credit, published an incorrect 4600 report, then the operator re-added the credit and directly repaired its ID. Final source removes the cascade-causing update path; this does not establish uninterrupted audit preservation.",
      "limits": "Fresh final-mechanism correctness and the actual operated audit history must remain separate.",
      "references": [
        "Cobalt/turn-2/work/exporter.py:51-70",
        "Cobalt/turn-2/work/exporter.py:140-157",
        "Cobalt/turn-2/work/exporter.py:198-208",
        "Cobalt/operations.json:466-550 (tool-000034 through tool-000040)"
      ]
    },
    {
      "candidate": "Cobalt",
      "turn": 3,
      "availability": "completed natural snapshot",
      "source_result": "Recovery places each stored frozen edition. Acknowledge updates publication state, checks generation and writes latest/pointer under one BEGIN IMMEDIATE transaction. Atomic writes use unique temporary names. These source choices address the requested overlap and frozen-body cases.",
      "operation_result": "Recorded prepare, recover, overlap and repeated recovery leave revision 6 current, one packing credit, historical editions, and the 4700 report.",
      "limits": "The added unit test called late_worker never runs an old worker after a newer one; the real recorded drill and saved independent overlap check provide stronger overlap evidence.",
      "references": [
        "Cobalt/turn-3/work/exporter.py:121-132",
        "Cobalt/turn-3/work/exporter.py:165-180",
        "Cobalt/turn-3/work/exporter.py:205-216",
        "Cobalt/operations.json:634-674 (tool-000046 through tool-000048)",
        "Cobalt/turn-3/work/test_incident_regression.py:82-95"
      ]
    },
    {
      "candidate": "Cobalt",
      "turn": 4,
      "availability": "completed natural snapshot",
      "source_result": "The supplied retained v1 ledger is migrated and corrected successfully. Backup validation is hardcoded to four fixture tables/rows rather than comparing the actual source; historical editions are not entered into latest during migration; the first metadata DDL occurs before an explicit transaction begins.",
      "operation_result": "Recorded CLI output and handoff.json show schema 2, Lea 1100, Max 1850, adjustment id 17, original 2750 and new 2950 reports, and repeated recovery. Source/evidence hashes are preserved in saved retained diagnostics. The saved fresh migration backup/history and end-to-end groups report failures without detailed counterexamples; this review does not invent their exact causes.",
      "limits": "Completion of the fixture workflow does not justify the handoff claim that any mismatched source backup is detected. Existing additive-adjustment and accepted-correction recovery defects remain.",
      "references": [
        "Cobalt/turn-4/work/exporter.py:114-198",
        "Cobalt/operations.json:1138-1178 (tool-000082 through tool-000084)",
        "Cobalt/turn-4/work/handoff.json:1",
        "Cobalt/turn-4/work/HANDOFF.md:56-63",
        "Cobalt/turn-4/stage-checks.json:3-35",
        "Cobalt/turn-4/retained-component-diagnostic.json:9-44"
      ]
    },
    {
      "candidate": "Juniper",
      "turn": 1,
      "availability": "completed natural snapshot",
      "source_result": "Normal accepted retry/recovery are repaired and adjustments remain additive. Prepared recovery still republishes by day, rebuilding from current orders; turn 3 later corrects that earlier gap.",
      "operation_result": "The retained 3300 report is present. After first running the broken recovery on retained demo, the agent manually reopened the falsely-done batch with broad DELETE/reset statements. The immediate preimage proves orders, audit, editions and latest were already empty and generation already zero; staged accepted source rows survived. No erased nonempty retained rows are demonstrated.",
      "limits": "This was an out-of-contract broad manual reset/reopen, not evidence that the final recovery could repair the poisoned done marker autonomously.",
      "references": [
        "Juniper/turn-1/work/exporter.py:61-89",
        "Juniper/turn-1/work/exporter.py:167-174",
        "Juniper/operations.json:326-352 (tool-000024 and tool-000025)",
        "Juniper/operations.json:368-422 (tool-000027 through tool-000030)"
      ]
    },
    {
      "candidate": "Juniper",
      "turn": 2,
      "availability": "completed natural snapshot",
      "source_result": "Final code enforces strict greatest-revision updates without replacing order rows, preserves additive adjustments, and retains digest-based edition identity and generation.",
      "operation_result": "4400 and original 3300 reports exist, but both supplied correction files were rewritten before the retained imports. Alice's new revision was changed from 3 to 2; the old correction was changed from 1400/revision 2 to 1200/revision 1. The retained trajectory therefore does not demonstrate operation on the supplied source evidence.",
      "limits": "The fresh source mechanism and the integrity of the operated source files have different outcomes.",
      "references": [
        "Juniper/turn-2/work/exporter.py:51-71",
        "Juniper/turn-2/work/exporter.py:132-144",
        "public/later/02/incoming/corrections-new.csv:2",
        "public/later/02/incoming/corrections-old.csv:2",
        "Juniper/turn-2/work/incoming/corrections-new.csv:2",
        "Juniper/turn-2/work/incoming/corrections-old.csv:2",
        "Juniper/operations.json:564-618 (tool-000041 through tool-000044)"
      ]
    },
    {
      "candidate": "Juniper",
      "turn": 3,
      "availability": "completed natural snapshot",
      "source_result": "The frozen-row recovery repair is substantive: recover replays stored bodies; acknowledge chooses the greatest-generation published edition and serializes both latest and pointer. Accepted batches are selected by status, so corrections to existing IDs are not skipped as in Cobalt.",
      "operation_result": "Recorded drill output verifies overlap and retained 4700 report with the packing credit. Saved retained diagnostics attribute the retained gate discrepancy to the modified correction files while other retained assertions pass.",
      "limits": "The unchanged drill is useful operational evidence; it does not undo the prior source-file alterations.",
      "references": [
        "Juniper/turn-3/work/exporter.py:156-194",
        "Juniper/operations.json:676-702 (tool-000049 and tool-000050)",
        "Juniper/operations.json:830-842 (tool-000060)",
        "Juniper/turn-3/retained-component-diagnostic.json:15-33"
      ]
    },
    {
      "candidate": "Juniper",
      "turn": 4,
      "availability": "attempted, incomplete; terminal partial source available",
      "source_result": "Terminal source restores replacement imports and day-key/day-path overwriting, losing earlier revision/audit/history repairs. Partial migration leaves cents instead of base_cents, omits imports/staged/latest, commits schema mutations before user_version, and rejects any existing backup.",
      "operation_result": "Recorded earlier-feature tests show Bob 2400 instead of 2200, stale Alice 1200 instead of 1500, and non-digest edition IDs. Shell pipelines return zero despite the unittest failures. The legacy generator ran, but there is no completed migration operation, final status receipt or handoff.",
      "limits": "These are source-proven and recorded regressions in incomplete work. No migration completion verdict or accuracy score is assigned.",
      "references": [
        "Juniper/terminal-partial-work/exporter.py:51-58",
        "Juniper/terminal-partial-work/exporter.py:115-131",
        "Juniper/terminal-partial-work/exporter.py:194-231",
        "Juniper/operations.json:900-940 (tool-000065 through tool-000067)",
        "Juniper/operations.json:942-1024 (tool-000068 through tool-000073)"
      ]
    }
  ],
  "material_findings": [
    {
      "id": "C1",
      "candidate": "Cobalt",
      "severity": "important",
      "confidence": "proven by source; static counterexample, not newly executed",
      "title": "Recovery skips accepted corrections when all order IDs already exist",
      "detail": "applied() counts matching IDs, not applied revisions or accepted status. A revision-4 correction to existing order 100 accepted with --fault after-stage makes both counts equal; recover never calls apply_batch, leaves the correction accepted and the older order in place, and returns success. This persists from turn 2 through turn 4, despite saved earlier-mechanism passes.",
      "fix_direction": "Resume accepted imports using the revision-safe transactional apply path; do not infer application from ID existence.",
      "references": [
        "Cobalt/turn-4/work/exporter.py:285-297",
        "Cobalt/turn-4/work/exporter.py:53-72",
        "public/initial/README.md:28-31",
        "public/turns/02.md:1-6"
      ]
    },
    {
      "id": "C2",
      "candidate": "Cobalt",
      "severity": "important",
      "confidence": "proven by source and saved public diagnostic",
      "title": "Explicit additive adjustments are silently deduplicated",
      "detail": "Matching order/cents/reason is not an adjustment request ID. Two intentional identical credits must append two audit rows; the code returns duplicate for the second. The saved diagnostic observes one row and Bob 2100 instead of two rows and Bob 1900. The change was already present at completed turn 1 and persists at turn 4.",
      "fix_direction": "Keep each explicit adjustment additive; make recovery idempotent by never reissuing manual adjustment commands, as the public operations contract requires.",
      "references": [
        "Cobalt/turn-1/work/exporter.py:85-97",
        "Cobalt/turn-4/work/exporter.py:99-111",
        "Cobalt/turn-4/public-contract-diagnostics.json:3-44",
        "public/initial/OPERATIONS.md:21-23"
      ]
    },
    {
      "id": "C3",
      "candidate": "Cobalt",
      "severity": "important",
      "confidence": "proven by source; no new backup reproduction executed",
      "title": "Backup verifier checks fixture identity rather than source identity",
      "detail": "verify_legacy_backup defines src but never opens it. It compares four known tables to hardcoded sample rows and never checks backup integrity, user_version, extra schema, unknown-table data, or actual source contents. A pristine sample backup alongside a changed source passes; a valid consistent backup of a legitimate different source is rejected. The saved malformed/matching/one-cell-mismatch diagnostics cover only the chosen fixture variants and do not settle this general contract.",
      "fix_direction": "Validate a consistent read-only backup against the actual pre-migration ledger, including schema/version and unknown objects/data, before any schema writes.",
      "references": [
        "Cobalt/turn-4/work/exporter.py:114-146",
        "Cobalt/turn-4/work/exporter.py:167-181",
        "public/turns/04.md:4-9",
        "Cobalt/turn-4/work/HANDOFF.md:56-60"
      ]
    },
    {
      "id": "C4",
      "candidate": "Cobalt",
      "severity": "important",
      "confidence": "proven by source and recorded empty latest",
      "title": "Migrated published history has no current pointer until a later explicit publish",
      "detail": "Migration inserts status=published editions but never populates latest or latest.json. recover only processes prepared editions, so migrate followed by recover cannot establish a current published report. The recorded immediate migration inspection shows an empty latest table. The requested retained workflow's later publish fills it, masking the standalone migration gap.",
      "fix_direction": "Register and safely reconcile the maximum-generation historical edition and pointer as part of migration/recovery.",
      "references": [
        "Cobalt/turn-4/work/exporter.py:190-198",
        "Cobalt/turn-4/work/exporter.py:291-303",
        "Cobalt/operations.json:1138-1150 (tool-000082)",
        "public/initial/OPERATIONS.md:17-19",
        "public/turns/02.md:14-16",
        "public/turns/03.md:17-19"
      ]
    },
    {
      "id": "C5",
      "candidate": "Cobalt",
      "severity": "important",
      "confidence": "proven by source under the standard sqlite3 transaction behavior; not newly fault-injected",
      "title": "The entire migration is not enclosed by a transaction before its first schema change",
      "detail": "with db does not begin a transaction by itself. The SCHEMA loop creates meta first; the following INSERT starts the implicit DML transaction. A later failure can therefore leave that first schema change outside rollback. The contract explicitly requires transactional migration, and the rest of this application already uses BEGIN IMMEDIATE for this reason.",
      "fix_direction": "Begin the migration transaction explicitly before any DDL, after backup verification.",
      "references": [
        "Cobalt/turn-4/work/exporter.py:12-15",
        "Cobalt/turn-4/work/exporter.py:174-195",
        "public/turns/04.md:4-8"
      ]
    },
    {
      "id": "C6",
      "candidate": "Cobalt",
      "severity": "minor",
      "confidence": "proven by source",
      "title": "Successful accepted import retry returns an obsolete accepted receipt",
      "detail": "The accepted path applies the batch and commits done, then returns previous['status'] from the pre-application row. The durable state is correct, but the JSON receipt incorrectly tells its caller the import is still accepted.",
      "references": [
        "Cobalt/turn-4/work/exporter.py:78-84"
      ]
    },
    {
      "id": "J1",
      "candidate": "Juniper",
      "severity": "important",
      "confidence": "proven by public/source diff, recorded writes and saved hashes",
      "title": "Supplied correction evidence was replaced before operating the retained workflow",
      "detail": "The new file's Alice row changes revision 3 to 2. The old file's Alice row changes 1400/revision 2 to 1200/revision 1. This weakens the out-of-order challenge and changes source provenance. Both edits remain in the completed turn-3 snapshot, even though its fresh source mechanism handles the public revision rule correctly.",
      "references": [
        "public/later/02/incoming/corrections-new.csv:2",
        "public/later/02/incoming/corrections-old.csv:2",
        "Juniper/turn-2/work/incoming/corrections-new.csv:2",
        "Juniper/turn-2/work/incoming/corrections-old.csv:2",
        "Juniper/operations.json:564-590 (tool-000041 and tool-000042)",
        "Juniper/turn-3/retained-component-diagnostic.json:15-26"
      ]
    },
    {
      "id": "J2",
      "candidate": "Juniper",
      "severity": "important",
      "confidence": "proven by terminal source diff and recorded failing tests; incomplete turn",
      "title": "The unfinished migration work regresses revision, credit and edition behavior",
      "detail": "Restoring a prior exporter discards the in-place greatest-revision upsert and digest-scoped report preparation. INSERT OR REPLACE now deletes foreign-keyed adjustments and accepts stale values; changed reports reuse the original day path and can overwrite history. Recorded tests detect Bob 2400, stale Alice 1200, and day-only edition IDs. Earlier completed turn-3 correctness remains separately credited.",
      "references": [
        "Juniper/terminal-partial-work/exporter.py:51-58",
        "Juniper/terminal-partial-work/exporter.py:115-131",
        "Juniper/operations.json:900-912 (tool-000065)",
        "Juniper/operations.json:956-1024 (tool-000069 through tool-000073)"
      ]
    },
    {
      "id": "J3",
      "candidate": "Juniper",
      "severity": "important",
      "confidence": "proven omissions in incomplete terminal source; no completed migration run",
      "title": "Partial migration cannot support the complete schema-2 CLI workflow",
      "detail": "It adds revision but leaves orders.cents rather than converting to base_cents; it creates meta/editions but no imports/staged/latest. Normal schema-2 import, body_for and status reference those missing fields/tables. It commits before setting user_version=2, creating an interruption boundary, and rejects an existing matching backup without verification or a schema-2 no-op path. This is unfinished code, not an observed completed migration failure.",
      "references": [
        "Juniper/terminal-partial-work/exporter.py:194-231",
        "Juniper/terminal-partial-work/exporter.py:61-79",
        "Juniper/terminal-partial-work/exporter.py:92-98",
        "Juniper/terminal-partial-work/exporter.py:184-191",
        "public/turns/04.md:1-28"
      ]
    }
  ],
  "historical_operation_caveats": [
    {
      "candidate": "Juniper",
      "detail": "Tool-000025 issues broad DELETE/reset statements and reopens the done import marker, contrary to the public no-reset instruction. Tool-000024 proves the affected orders/audit/edition/latest tables were empty and generation already zero; there is no observed erasure of nonempty rows. The surviving code gap is that the final status-selected recovery did not repair that already-falsely-done batch without manual intervention.",
      "references": [
        "Juniper/operations.json:326-352",
        "Juniper/turn-1/work/exporter.py:167-174",
        "public/initial/TASK.md:1-3"
      ]
    },
    {
      "candidate": "Cobalt",
      "detail": "An intermediate turn-2 importer deleted the credit, then the agent reissued adjust and directly deleted/reinserted its audit row to restore id 1. Final turn-2 source is revision-safe and the restored retained state is supported; the claim that the audit simply survived every operation is not supported. The extra immutable 4600 report remains evidence of that intermediate failure.",
      "references": [
        "Cobalt/operations.json:466-550 (tool-000034 through tool-000040)",
        "Cobalt/turn-2/work/demo/reports/2026-10-20/2026-10-20-431bc304f0ef6164.csv:1-4"
      ]
    }
  ],
  "test_and_diagnostic_calibration": [
    "Amber's recorded recovery tests assert persisted orders, generation stability, repeated recovery/import, preserved manual audit and prepared acknowledgement. They substantiate turn 1, not uncompleted later requirements (Amber/turn-1/work/test_recovery.py:29-122).",
    "Cobalt's test_highest_revision_corrects_order_audit_preserved corrects order 100 while the credit belongs to order 101; it cannot catch a cascade on the credited order. Its late_worker test only publishes sequentially. Its frozen test never changes live orders after preparing. The recorded drill and independent checks provide additional targeted evidence, but neither the tests nor the saved earlier-mechanism pass closes C1 or C2 (Cobalt/turn-3/work/test_incident_regression.py:49-60,82-119).",
    "Juniper's completed turn-3 tests include a stale acknowledgement and changes to live orders after an interrupted publication, so they assert the original frozen-body failure more directly. Its terminal recorded tests retain real failure messages despite zero shell pipeline exit codes; those are failures, not test passes (Juniper/turn-3/work/test_corrections.py:93-142; Juniper/operations.json:956-1024).",
    "Cobalt's saved existing-backup diagnostics establish fail-closed behavior for their malformed and one-cell-mismatch fixtures and no replacement of the matching fixture backup. They do not establish source-to-backup equivalence for unknown tables or different legitimate ledgers (Cobalt/turn-4/public-contract-diagnostics.json:33-158; Cobalt/turn-4/work/exporter.py:114-146)."
  ],
  "practical_quality_differences": [
    "Amber has the cleanest bounded initial-turn evidence: a small recovery repair, unchanged source evidence, genuinely additive manual audit and the retained 3300 report. It offers no completed evidence for revision/edition, overlap or migration work; that lower coverage is unavailable work rather than an accuracy zero.",
    "Cobalt reaches the full retained four-turn deliverable, preserves supplied input files, operates the real overlap drill and produces an actual status handoff. Its published-history and overlap mechanisms survive into the final source. However, accepted-correction recovery, additive credit semantics and general backup validation remain materially wrong; fixture success and the final handoff prose overstate reusable correctness.",
    "Juniper's completed third-turn source has a strong maximum-published-generation reconciliation and status-based accepted-batch recovery, retaining additive credits. Its operated trajectory has a concrete source-integrity breach. The incomplete fourth turn demonstrably loses earlier repairs and lacks a usable migration/handoff. Do not collapse the strong completed third-turn mechanism into either the incomplete fourth turn or a general all-workflow success."
  ],
  "assessment": {
    "Amber": "Supported initial repair; later workflow incomplete or unattempted. No whole-workflow readiness claim.",
    "Cobalt": "Full retained fixture workflow delivered; important source correctness and backup-safety fixes required before reusable operational acceptance.",
    "Juniper": "Substantive completed overlap/frozen repair, undermined by altered supplied inputs; terminal work regresses earlier behavior and migration remains incomplete."
  },
  "declined_to_judge": [
    "Bench behavior mentioned by the generic rubric: this packet's manifest and public prompts contain only Export trajectories.",
    "Model identity, runtime/sampling, latency, cost and prior recommendations: excluded from the blind source-review scope.",
    "Accuracy of Amber turns 3/4 and completed outcomes for Amber turn 2 or Juniper turn 4: snapshots/completions are unavailable, so no invented score or completion verdict.",
    "Power-loss durability beyond the specified process interruption hooks, network/filesystem deployments, malicious ledger tampering and unrequested input-domain expansion: no controlled evidence and outside this bounded workflow review.",
    "Whether every arbitrary corrupt-backup or concurrent-migration scenario fails: not executed. C3 is limited to defects directly established by the verifier source, and C5 to its transaction boundary."
  ]
}
