[HMarcusWH]

What this repairs

The required archive-reproducibility check has failed with E_TOOLING_CLOSURE ("Complete canonical code/guide closure differs between T and C") on every head since two reviewed integrity files changed after the historical tooling snapshot T 71cc34fa:

• verification/integrity/workflow_contract.py
• verification/integrity/tests/test_workflow.py

The v0.2.0 verifier requires the current verification/integrity/, verification/archive/ and guide closure to equal T's. That fails on main, on PR #115, and on PR #113's own final head 117f4905 (run 40778332081). The archive is not damaged. At the accepted base B 17c315f7, the full check including committed publication passed (run 39513915824).

The staged 30-workflow T2 candidate cannot be switched on as a fix. Its inherited verifier needs all twelve required checks of #113's final head to have succeeded, but archive-reproducibility failed there. It also needs an APPROVED review from an owner-authorized roster, but #113 has only COMMENTED bot reviews and the roster is UNCONFIGURED. This PR is the narrow route the owner chose: repair the archive check for the existing 25 workflows, approved under a transparently named owner engineering policy.

How the new gate works

The archive job's last step now runs verification/research/archive_successor_gate_v1.py instead of the candidate-bound equality check. The archive adversaries still run first, unchanged.

1. Historical T, verified by T's own code. A fresh, clean clone at T runs T's unchanged python -B -m verification.archive.offline_smoke:

   • in standalone mode, on the committed supplement;
   • with T's own nonroot unshare --net/setpriv isolation;
   • with no GitHub, Git, token or import variables.

   The gate accepts only an isolated, T-bound, accepted-artifact-scope PASS with the original stage list. I checked those expectations against the real report from B's passing run. Standalone mode compares the packaged tooling with the code that executes, so a pass also proves T's code ran, not C's. A blocked replay stays BLOCKED.

2. Current C, checked as Git objects. C must be the checkout HEAD, with no modified tracked files. pr113_tooling_migration.audit (reused, not reimplemented) checks:

   • T and B are ancestors of C;
   • the supplement is unchanged since B;
   • the 35 packaged T files are byte-exact;
   • only the two named closure files differ;
   • there are 25 workflows.

   The gate adds stricter checks:

   • the whole archival/v0.2.0 tree equals B's;
   • each changed closure file is at exactly its approved blob;
   • research-integrity.yml is exactly the approved change of its preactivation blob;
   • C's own run_tests --suite workflow passes with no skips.

   The source audit is repeated after both legs.

3. Approval. archive_successor_policy_v1.json has authority OWNER_ENGINEERING_POLICY, owner HMarcusWH and independent_security_review: false. The gate rejects true. The policy takes effect only through the owner's merge. It is not an independent security review, not a PR WIP: finish mathematical implementation and verification infrastructure (certify in successor PR) #113 approval, not certification and not release acceptance. The PR WIP: finish mathematical implementation and verification infrastructure (certify in successor PR) #113 final-head failure stays recorded as a failure, and the gate does not depend on it.

Any third closure change, or a later edit of an approved file, fails with E_SUCCESSOR_DELTA until the policy is changed and merged again.

Files

Path	Change
.github/workflows/research-integrity.yml	archive step runs the gate; evidence directory added to the existing upload
verification/integrity/workflow_contract.py	archive command, required-verifier list and APPROVED_WORKFLOW_SHA256
verification/research/archive_successor_gate_v1.py, archive_successor_policy_v1.json, test_archive_successor_gate_v1.py	new gate, policy and 19 tests
verification/research/archive_successor_approval_v1.py	stdlib helper that parses the single approved workflow change
verification/research/pr113_a2_successor_identity.py	the two pins for the changed files
pr113_workflow_profile_v2.py, pr113_successor_workflow_profile.py, pr113_atomic_migration_preview.py and three of their tests	the inactive 30-workflow candidate's "original 25 unchanged" checks accept exactly the owner-approved change, read from the policy at the inspected head, and still reject anything else
docs/operations/ARCHIVE_SUCCESSOR_CONTRACT_V1.md	runbook

Unchanged: the historical bundle, verification/archive/, every recorded hash, the packaged T snapshot, verify_bundle.py (nothing relaxed), the workflow count (25), and the staged candidate's manifests, sources and review packet. No test is skipped or removed.

Evidence at 9054a48

• Gate tests: 19/19. Each covers a rejection path: policy mutations, stale or foreign approvals, report forgeries (wrong binding, missing isolation, skipped or added stages, candidate-CI scope), blocked and failed legs, a mutated T checkout, source drift, output overlap, environment scrubbing and the process-group timeout kill.
• Integrity suites: workflow 8/8, context, reference, preservation and tex pass. In archive, only the real-namespace case fails locally, identically on main, because it needs a nonroot runner.
• All 257 existing research test modules give the same result on main and on this branch. This was run without Lean in the worktree; the Lean-dependent modules fail identically on both.
• Running the gate CLI on this repository passes its policy and source audit, then reports BLOCKED as root, which T's launcher requires.
• Genuine isolated T replay, in this PR's own CI (Research integrity run 41207232823): archive-reproducibility passed for the first time since the tooling changed.
  • T's own verifier ran in standalone mode as uid 1001, with network isolation ENFORCED and all capabilities zero.
  • It completed all 12 original stages, including the legacy 29-test replay ("Ran 29 tests … OK"), "FINAL kit verification PASS: 21 pages, 17 claims, 340 protected canonical blobs", and the C10 consistency check with its audited axioms.
  • The gate passed all 8 of its stages. The current-C workflow suite ran 8 methods and passed. Approval is recorded as OWNER_ENGINEERING_POLICY, independent_security_review: false.
• Every other check on 9054a48 is green: Research Lean (the full research suite, axiom audit, C10, D0-F, default target, root imports, placeholder rejection), Lean, Executable verification, Application contract on Ubuntu and Windows, Integrity portability, frozen-source preservation and release-export integrity.

The owner has instructed that this PR be merged once it is green. That merge is the owner-engineering approval event recorded in the policy.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KdhJ3TNPDaXGbNLXiKuZPC

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-31T14:45:17.498732Z	9054a48	Draft marked ready
🔒 Security Review	✅ Completed 2026-10-31T14:48:38.450978Z	9054a48	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.