docs(007): publish audit — keep seed-morph78 weights local-only

[scrimshawlife-ctrl]

Recommendation

Do not publish the seed-morph78 weights yet. Full audit: specs/007-hyperlexical-model/PUBLISH-AUDIT-20260924.md.

Blockers (OBSERVED, evidence in receipts/publish-audit-20260924/)

1. Promotion metric is train-contaminated. 193 of the 256 soft_ceiling broad-val rows are in force_train_morph78_expanded.jsonl. On the 63 leak-free val rows, morph78 and morph65 both score 1.0, so the whole 0.9883 vs 0.8867 margin comes from trained-on rows. The 440-row test split has never been scored.
2. Not reproducible from main. Training ran from Spark ~/Hyperlex at 78c4d83 (39 behind main) with a 799-line uncommitted diff in export.py / loop.py / unbind_recipe.py, plus 86 settle/finish/eval scripts in ~/hlx outside git. The train receipt has no code commit.

Other gaps: data rights for crawled slang + 14 CC BY-SA rows; license conflict (MIT / Proprietary / apache-2.0 card); classify overconfidence; A1 150M reading; no Hub loader; card has no test metrics.

The audit lists a 7-step fix order, ending with a private/gated Hub repo first.

Not changed

name_gate, BEST pin, receipts, code. Read-only eval on Spark (no training). Evidence copies of off-git scripts are included for review.

Validation

pytest 747 passed; mkdocs strict PASS; evidence scanned for secrets (none).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9ac1437d1d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include the leak-free evaluation input

Even with the local checkpoints and dependencies installed, this audit cannot be reproduced from the committed evidence because audit-clean-val.json is read from an absolute off-repo path, while neither that file nor code that derives its 63 rows is included anywhere in the repository. Consequently, reviewers cannot verify the row exclusions, the claimed 193/256 overlap, or the reported 1.0 tie that supports the central B1 conclusion; commit the evaluation rows or deterministically construct them from committed inputs.

Useful? React with 👍 / 👎.