Skip to content

cbz_drift: reviewed drift acknowledgements; acknowledge the CapFirstExit extraction - #753

Closed
d6g8k5htny-coder wants to merge 1 commit into
mainfrom
claude/cbz0-drift-ack-20261105
Closed

d6g8k5htny-coder wants to merge 1 commit into
mainfrom
claude/cbz0-drift-ack-20261105

Conversation

@d6g8k5htny-coder

@d6g8k5htny-coder d6g8k5htny-coder commented Nov 5, 2026 •

Copy link
Copy Markdown
Owner

Change and purpose

federation-drift has been red on main since 2026-11-02 in tips mode: the scheduled runs failed on 11-02, 11-03, 11-04 and 11-05, and every PR's tips crawl inherits the failure (#730 at 5d6efa81). Triage: #738/6669319293.

The one block. CapFirstExit.frontier_image_subset has two copies in Math- (f532c3e2):

Copy Context Statement
frontiers/cap_first_exit_lean_20261002/CapFirstExit.lean:359 open Set unitInterval the same in both
frontiers/source_derivative_gaussian_20261025/inputs/elder/ReusedCompactFrontier.lean:19 open Set the same in both

The copy is a deliberate single-theorem extraction. It imports only Mathlib.Topology.Separation.Hausdorff, which does not provide unitInterval. Its binding record, COMPACT_FRONTIER_INPUT_BINDING.json, says: "Dropped unused open unitInterval only; open Set, namespace and implicit variables retained exactly", and a strict compile receipt backs it.

Why not a generic fix:

What this adds. An explicit, reviewed acknowledgement file, tools/cbz_drift/acknowledged_drift.json.

  • Each entry names one declaration, the exact set of its interface hashes (SHA-256 of kind\0interface), a reason and evidence.
  • A drift whose variants are exactly that set is reported as INFO LEAN-DRIFT-ACK, carrying the reason.
  • Any other change to either copy no longer matches and blocks again.
  • A malformed file is MALFORMED (BLOCK).
  • Every drift finding now reports its variants' hashes, so a new acknowledgement can be written from the report itself.

This PR adds one entry, the CapFirstExit extraction above. Scope: engineering (the drift crawler and its tests). No workflow file changes.

Current disposition

Verification

  • Drift modules. Run from tests/ as the workflow does: test_cbz_drift, test_cbz_drift_adversarial_01a103c8, test_cbz_drift_canonical_input and test_cbz_drift_pr_trigger_01a11c48 give Ran 83 tests, OK, in normal mode and with -O.
  • Mutation harness. It includes three new designated mutants: ack matched by name only, ack ignored and malformed ack accepted. The existing drift downgraded mutant is re-anchored to the new line. All are killed by assertion.
  • New tests:
    • an acknowledged drift is INFO, and a further change blocks again;
    • an acknowledgement for another name does not apply;
    • three malformed files each give MALFORMED.
  • Real federation (seeded from ULW 202f9f5c, main overridden to this head):

Hosted checks at eea5104a: all 10 green. That covers the required verify, public-intake and public-shop, plus federation-drift (pins and tips: its first green run on a current main base since 2026-11-02), drift-crawler-self-test, architecture, navigation, formal-evidence, landing-checks and browser-smoke. The merge state is clean.

Evidence and limits

  • What an acknowledgement records. That a reviewer has accepted one exact pair of interfaces. It does not establish that the two elaborate to the same term (B-006), and the entry's reason says so.
  • What it gives up. One more file to review: every new entry is a review decision. Entries are keyed by exact interface hashes, so they cannot silently cover later changes.
  • Authorship: 01DKKe (Anthropic Claude, session session_01DKKe26v3CSgdYch8kiJLt9). Credit 0. Scientific effect NONE.

https://claude.ai/code/session_01DKKe26v3CSgdYch8kiJLt9


Generated by Claude Code

…xit extraction

federation-drift has been red on main since 2026-11-02 in tips mode with one
LEAN-DRIFT: CapFirstExit.frontier_image_subset in Math- has two copies whose
statements are byte-identical and whose context differs only by an
`open unitInterval` that the reused copy drops on purpose. The copy imports
only Mathlib.Topology.Separation.Hausdorff, and its binding record
(COMPACT_FRONTIER_INPUT_BINDING.json, successor_context_delta) states the
change.

Neither generic fix is sound: editing the hash-bound copy breaks its
evidence and narrow import, and ignoring extra opens would miss real drift
(open Real vs open NNReal). This adds an explicit, reviewed acknowledgement:
tools/cbz_drift/acknowledged_drift.json lists a declaration and the exact
set of its interface hashes (SHA-256 of kind\0interface). A drift whose
variants are exactly that set is INFO LEAN-DRIFT-ACK with its reason; any
other change to either copy no longer matches and blocks again; a
malformed file is MALFORMED (BLOCK). Each finding now carries its
variants' hashes.

Three tests and three designated mutants cover it, and the existing
drift-downgrade mutant is re-anchored to the new line. The four drift
modules run 83 tests OK in normal and -O mode. Crawling the federation:
pins 0 BLOCK; tips 1 BLOCK becomes 0 BLOCK with one LEAN-DRIFT-ACK.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DKKe26v3CSgdYch8kiJLt9

Copy link
Copy Markdown
Owner Author

Closing as a duplicate of #751. 01DKKe, the author.

#751 fixes the same federation-drift BLOCK with the same kind of mechanism, and it came first:

It is also stricter. It keys on the exact location set (repository, path and interface hash), and it reports stale entries as EXEMPT-UNUSED. #753 keys on the name and the hash set only.

My triage (738/6669319293) missed #751, and that is my error. Nothing in #753 is needed on top of #751. A refreshed-base check of #751 against current main is at 751/6670566736: 0 BLOCK at tips and at pins. The branch claude/cbz0-drift-ack-20261105 stays in place, so this PR can be reopened if reviewers want it.


Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants