ANN07 actual retention audit, 2026-10-02
Disposition: verified failure of exposed-label retention for all five fitted candidates. Each candidate fails the fixed .02 maximum raw-accuracy-loss gate on both A and B. C passes. Exact frozen router replay on all7,020 original common public states passes with zero probability, confidence, index or action disagreement. Full ANN07 and deployment remain unaccepted; low-budget backend outcomes are pending.
Original identities and method
The independently selected original receipt byte SHA is 76da38efeeb801609b48295fe135f7c8d17b3de245b6549d4a0dcb277c304a24, canonical identity 262576c4e353e719418fe46a663beea3be6795455401742c83f78e3de8864ff7. Original PLAN ffa0849d686c9caa6e88f1f244c3854153212448b4ef587b08aedd5cff50464c and FIT manifest 195dfbeffb8fc872eeb230f910c7429403886586c62e89a8b8c796fa077b8e2f remain unchanged. The external measured JSON pins every incumbent/candidate/router envelope and original anchor/panel/replay reference.
This reviewer authored a separate metric/gate aggregation loop; it did not call the producer retention-evaluation or replay wrapper. All raw/masked losses, role/class accuracies, confusion matrices, mask-change counts and per-label drops exactly matched retained outputs. The frozen shared encoder/numerical core and task semantics were reused, so this is not an independent numerical reimplementation. Router reference forward/encoder/mask/decoder code was independently authored by the security lane and reviewed by this lane; it was reused directly and compared against both the retained per-row replay and the frozen harness.
All150 original exposed examples exactly equal the three original byte-pinned ANN01 corpora in order. Labels were rechecked against the frozen teacher and legal masks. Every original protected context from all60 train/30 selection episode receipts was freshly pinned, catalogued and ordered to reconstruct the full7,020-state panel. All state digests and record IDs match; there are7,020 unique record IDs. All five candidate envelopes imported with the original approved content digest. Exports of every candidate, incumbent and router remained unchanged. All24 protected snapshot leaves and the distinct11 executable source identity were checked;136 original source/data files were freshly rehashed after inference with zero mismatches.
Failure reproduced
The incumbent raw and masked accuracies are A94.6667%, B100%, C100%. Raw and masked candidate metrics are identical here; masks do not rescue the failures. The table shows percentage-point drops from incumbent.
| Initialization | A raw/masked | A drop | B raw/masked | B drop | C raw/masked | Gate |
|---|---|---|---|---|---|---|
| 431 | 51.3333% | 43.3333 | 90.0000% | 10.0000 | 100% | FAIL A/B |
| 977 | 45.3333% | 49.3333 | 83.3333% | 16.6667 | 100% | FAIL A/B |
| 1543 | 52.0000% | 42.6667 | 90.0000% | 10.0000 | 100% | FAIL A/B |
| 2203 | 47.3333% | 47.3333 | 83.3333% | 16.6667 | 100% | FAIL A/B |
| 3251 | 49.3333% | 45.3333 | 83.3333% | 16.6667 | 100% | FAIL A/B |
The preselected431 therefore fails: A51.3333% and B90%, with losses43.3333 and10 percentage points. It cannot be replaced with another seed to conceal the failure; every other seed also fails. The unchanged threshold is .02 with explicitly local1e-12 tolerance, not the historical frozen helper's unqualified floating-point comparison.
Class supports, in wait/alpha/beta/gamma order, are A[29,46,38,37], B[125,10,8,7], C[109,13,13,15]. Incumbent raw correct counts are A[26,46,37,33], B[125,10,8,7], C[109,13,13,15]. Candidate class correct counts follow; exact raw/masked per-label drops and4x4 confusion matrices are retained in the measured JSON.
| Initialization | A class correct | B class correct | C class correct |
|---|---|---|---|
| 431 | [0, 46, 26, 5] | [125, 10, 0, 0] | [109, 13, 13, 15] |
| 977 | [0, 46, 22, 0] | [125, 0, 0, 0] | [109, 13, 13, 15] |
| 1543 | [0, 46, 27, 5] | [125, 10, 0, 0] | [109, 13, 13, 15] |
| 2203 | [0, 46, 25, 0] | [125, 0, 0, 0] | [109, 13, 13, 15] |
| 3251 | [0, 46, 28, 0] | [125, 0, 0, 0] | [109, 13, 13, 15] |
C retains all41 nonwait anchors as well as109 waits. B431/1543 retain alpha restart10/10 while losing beta8/8 and gamma7/7; the other seeds miss all25 restart positives. This explains why B90% for431/1543 still fails against the100% incumbent. A has broad exposed-anchor regressions beyond the gamma selection regression already disclosed. These actual negative results establish that the current supervised update does not satisfy retention.
Scope, chronology and immutable proof copy
The reviewed runner constructs and exclusively persists ADMISSION and the complete original panel before its inference calls. Their exact hashes/selection and current source order were checked. This is source and retained-file evidence of cooperative execution ordering; it is not an external sealed timestamp or hostile same-user attestation. No second producer evaluation was run. The reviewer ran only pure inference, without optimizer updates, fitting, HTTP, workers, protected release outcomes or deployment.
Frozen proof manifest 1813545a1fe016f9da97bad6f804f64dbaed8470ba3d2da67c694afea69eb49a was freshly verified for all25 leaf hashes, sizes and exact file set (49474572 bytes). Its historical inputs/new evaluator source/output copies are verified artifacts; protected source remains the separately pinned unchanged physical snapshot. Independent package readback SHA 1b27caa0f4d60fb3705e21f686c66aaa2655883af6dc22aefc0fafd25322c95a.
The failure is a measured model-quality result, not a reason to revise the fixed gate after exposure. Old anchors are previously exposed supervised examples; router replay is unchanged source/weights behavior on already-exposed states. Neither result proves fresh generalization, real repair effects or reward learning. Pending low-budget outcomes and later architecture/recipe remedies require separately reviewed/predeclared successor work; there is no automatic promotion or activation.
Measured actual audit (ANN-07-RETENTION-ACTUAL-AUDIT-2026-10-02.json), SHA a10008bcedfe18e2f831e2f0eb09dd90a2c271511df976a688df6fd8e27bd1e8. External audit instrument (ANN-07-RETENTION-AUDIT-2026-10-02.mjs). Prior source-only acceptance remains in ANN07 source review (ANN-07-INFERENCE-SOURCE-REVIEW-2026-10-02.md) and does not imply model-quality acceptance.