12 lines
1.8 KiB
Markdown
12 lines
1.8 KiB
Markdown
# E1b target-access metadata erratum and independent audit
|
||
|
||
MeltAssist independently audited68 checks and returned PASS_WITH_METADATA_ERRATUM; Claude1794908035 accepted. The original E1b report, scoring JSON, source, target-access logs and local proofs remain unchanged. No recomputation is performed.
|
||
|
||
The reused training helper hardcodes `role=fold0_learning` in target-access logs even for fresh fits in folds1–4. This is a metadata error. The auditor checked every actual case list and target hash against the corresponding frozen learning role: fold1 374, fold2 366, fold3 291 and fold4 350 parts. No wrong-fold substitution or test leakage was found. Use the enclosing fold and actual case/hash identity, not that hardcoded role string, to interpret the retained logs. Any new implementation should record its actual fold without mutating the historical source.
|
||
|
||
The audit checked552 prediction hashes,552 feature hashes,552 A-file hashes,13 local freeze proofs and the Slack review ZIP. For42 invalid-G parts,36 retained explicit finite-count0 metadata; for the other6, the evidence was the frozen SCIENTIFIC_MISSING source path plus file hashes, not a fresh numeric scan. The audit is local/read-only and is not an independent remote-publication acceptance.
|
||
|
||
Interpretation remains: pooled field confirmation, secondary τ interval positive, last-fill improvement unconfirmed/difference uncertain. A CI crossing0 does not establish unchanged or equivalent performance. Five-fold pooled confirmation does not mean every fold improves.
|
||
|
||
Exact received files: `artifacts/meltassist-e1b-audit-2026-11-17/audit.json`, SHA256ba386fed3ec97ccc0d82a27e084bce0c39d31565e97a786c941d092cb5b0e730; `audit.py`, SHA256381ebceb25e5209595b19ad91d307f5e170a876f8517523649bd3d39a7d61261. Melt independently verified those two byte hashes before queuing them for the existing single-writer publication.
|