[realmarcin]

The D52 evaluator, the E3x evaluation routes it is built on, and the fixes from claude's independent review of those routes, published from a history-clean branch.

What changes

The D52 scoring rules, a new evaluator instrument (b38117d9…):

• the owner's D52 rules, including D61's scope: the two committee-contact locations count as one criterion wherever governance_committee_contact is scored, rubric10 E2.2 included;
• historical scores are not rewritten;
• re-scoring all 566 committed records changes no score, maximum, item or evidence pointer. Only CM4AI's 134 Q9 coverage labels change, and they now name the regulated-access criteria D52 excludes.

The E3x evaluation routes, 8 commits from d6eed9274:

• identity-aware blinding, shared with registered evaluation;
• deterministic reviewer copies, with blind keys kept outside reviewer directories;
• Unicode-safe short blind identifiers;
• arm-comparison checks that are invariant under source revisions;
• ingestion of registered evaluations into the comparability audit.

The E3x review fixes, 14 commits, F1–F12:

• F1: real provenance arms join registrations through declared mappings, and declared authority roots let the audit read real registrations.
• F2: retained attempts are anchored to receipts and execution locks.
• F3: declared comparison identities (arm, condition, method) are blinded.
• F4: structure and size fingerprints are disclosed in the private key, never in reviewer bytes, and nothing scientific is redacted.
• F5: the audit roster and the review cohort are declared data (configs/evaluation/*.example.yaml, docs/evaluation-cohorts.md), with no hard-coded project table.
• F10: tests kill every guard-removal mutant.
• F6–F9, F11 and F12 are also fixed.

#5125

A provenance record from the real writer now passes prepare, and a writer-shaped round trip retires its tuple from the remaining-rating queue. #5125 can close when this PR merges.

History

This branch publishes ee4a3547d without the commits that added two 1.5 MB file-inventory tarballs:

• the final tree of the publication commit fb34ed882 is identical to ee4a3547d's;
• the compact evidence in the tree is byte-identical to the tarballs' members, except the two 2.98 MB file inventories, which are recorded by hash;
• the largest blob in the published history is api_runner.py, at 441,846 bytes.

claude checked the tree equality and the blob history independently.

Reviews

• The D52 evaluator: claude's reviews of 4cd8f2a51 and ee4a3547d found it merge-ready. D1 and D3 are resolved, and D2 was settled by D61.

• The E3x routes: claude's independent review (e3x_d6eed9274_claude_review_2026-10-29.md) found 12 findings. Its re-review of the fixes (e3x_fixes_9f6d1583f_claude_rereview_2026-10-29.md) is merge-ready:

  • F1–F12 are resolved;
  • 649/649 pass on the head and on a scratch merge into main;
  • all guard mutants are killed;
  • a new project declared only in a manifest works with no code change.

• Five small follow-ups (R1–R4 and R6) are filed as issues, and are being fixed on this branch:

  • an audit crash with --authority-root;
  • advisory lock anchoring;
  • six untested new guards;
  • a generic refusal reason;
  • substring matching of the replicate token.

  R5 and R7 are watch items for before the first billed rating.

Tests

649 passed on 9f6d1583f: the 597-test evaluator and E3x set, plus the fixes' tests.

Not included

No rating runs. #5126 (the v8 comparator's rating route), #2929 and #3343 stay open, with plans and estimates; each needs its own billed run and the owner's word.

🤖 Generated with Claude Code