Independent byte and reference QA of review 61

The assessment files were left unchanged while their stored links and hashes were independently checked. There are 975 records: 615 existing review entries from 59 and 360 supplementary observation-field and explicit-negative-boundary assessments of the same candidates. This does not add 975 independent requests or training answers.

I authored the preparation 61 reference generator, but did not author these actual assessments. This is a storage, byte and reference check independent of assessment authorship. Earlier source and metadata exposure is disclosed: the check is not blind and establishes neither independent data origin nor renewed natural-language semantic accuracy.

The fixed manifest SHA is f8f0b2320bb214f102575384cace477067a0f075b98ee09c071e778d9285224b; the resulting receipt is report.json. Checks cover the hashes and sizes of 31 JSON artifacts, four original Go files, 98 raw spans, 105 declaration links, and 49 root plus 172 component relations. Multiple IDs sharing an entire enum declaration retain their stored relation order. Historical formatted, normalized and bundle hashes were compared through stored pointers. No Go AST, source formatter or candidate function was executed.

All 975 record digests, 975 whole quotes, nine material-clause quote ranges, 1,860 request/caption references and 6,680 evidence links matched. The 615 original and 360 supplementary entries have exhaustive one-to-one shard/manifest mappings. The 60 parents, 180 candidate positions, 15 contracts and candidate-count histogram of 11 parents with two candidates, 38 with three and 11 with four were preserved. A final reread confirmed that all 31 inputs and four original sources remained unchanged.

Existing entries	Scoped consistent	Contradicts	Omits	Uncertain
All 615	446	94	10	65
Candidate source fidelity	177	0	0	3
Candidate request coverage	32	94	7	47

These are recomputed counts of the author's assessment states, not new judgments. Source fidelity and request satisfaction are separate axes; a consistent finding is not a correct-candidate label. The selected 18 unknown parents, global 21 unknown parents, 17 total groups and 16 groups containing known outcomes remain unchanged. The earlier 60 overlay SHA 1beb5c06d3fb70c0a5e0727de793f1890b4ab4a47f1d392a6666b068f096c526 also remained unchanged. Assessments, group IDs, source IDs and historical truth are provenance, not scorer features.

There were two actual metadata QA invocations. The first failed with array_shape because the verifier assumed legacy Components:null was an array. This was a verifier assumption about the stored absence representation, not a discovered assessment defect. Its failure report was preserved. After narrowly supporting that representation, the second invocation passed. Three synthetic race runs passed; the final run confirmed 32 named test/subtest passes and the package pass. All three vet runs passed. The history is in ledger.json.

The writer's one successful assessment generation, two preparation builds with one failure, and two reader byte-QA attempts with one failure are separate histories and are not mixed with this verifier failure. Assessment mutation/regeneration, new semantic judgments/truth/labels, original graph/seed/role assignments, fitting, weights, separate model or candidate/API runs, protected-final reads, shared-repository writes and publication remain zero. Zero separate judge processes does not imply zero usage or cost for this AI-assisted Codex collaboration; that cost was not measured. training_ready remains false.