Repository navigation
Audit silent omission of HPE Lean axiom checks and provide an inventory repair - #1608
Conversation
|
FINDING / real-toolchain compatibility of the PR1608 repair: passes on the pinned HPE sources. This follows the PR's open full-build question and scoped claim. The actual Lean build-and-probe pipeline preserves all 42 named axiom reports after applying the exact proposed patch. No subprocess results or compiler reports were substituted. Exact inputs and environment
The three local source packages were copied into a new temporary tree without local build artifacts. Observed controls
All four rejected cases fail with Reproduction of the full-build comparison Use a fresh isolated checkout/worktree at the base above, with Lean 4.21.0 available. Fetch PR1608, extract its exact # From research/hyperplane-end-lean in the isolated checkout:
lake exe cache get
python3 -B verify.py --no-cache
# From the isolated repository root:
git apply /absolute/path/to/repair.patch
# Return to research/hyperplane-end-lean:
python3 -B verify.py --no-cacheFor the controls, change only Actual repaired canonical driver outputScope: this supplies the missing real-build compatibility evidence for this environment. It does not add a theorem, run Credit Jeremy Kothe/jeremykothe2 (PR593), DreamingOfClouds (PR599 and LemmaA/G), the PR1608 repair, and WeberChen8128/PR893's earlier coverage principle. Prepared for Hanson Sun (@bbMisk) with substantial OpenAI Codex assistance. This is a separate actual execution, not independent human peer review. Final overlap refresh: 740 open PRs and 267 discussion records, complete pagination and no warnings; titles/bodies/local notes do not cover every diff. |
|
Batch review at Merged to Reviewed scope: The standalone HPE driver can silently accept an empty derived probe inventory; controlled transcripts reproduce the gap and test an explicit-inventory repair. Existing intact HPE evidence is not refuted, and full repaired HPE/Mathlib adoption remains a separate task. The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions. Batch review, attribution and evidence. Thank you for the contribution. |
2 participants
Change and mathematical scope
The standalone HPE verifier can report
PASSafter silently losing axiom-check coverage. Removing every query makes it report 0 results. Indenting all 42 queries has the same effect, although Lean 4.21 accepts the indentation and emits their reports.This adds a source-bound audit and a proposed inventory repair, entirely in
research/hpe-axiom-inventory-audit/. The patch requires exactly the intended 42 theorem names before building, then one allowed-axiom report per name. It rejects missing, duplicate and unexpected reports. The frozen driver and Lean sources are unchanged.Scope: verification tooling. Driver experiments substitute subprocess results; they are not HPE compiler logs or proof replays. A separate real Lean 4.21 run verifies the small syntax fixture. No invalid accepted HPE theorem, new formal lemma, supplier improvement or new kappa is established.
Review: README and limits, proposed repair, deterministic evidence.
Task and overlap
Initial main:
ebedc0895eb4c06725af020bc1af165722b0a3ce. Submission base:81e81af67714bdc8d1693b0bf5c29468b92f62ff; all 21 consumed input files are unchanged. Taskformal-interface-lemma, scope keye3ffea67887eeb4c396c, changed to the verification-coverage prerequisite after overlap review. Revision 3 was read at refresh; the recorded assignment retains revision-2 provenance. Exact inputs, selection and scope are in TASK/SOURCES.Discussion1593 tracks the claim. The final focused search covered 733 open PRs and 244 Discussion records with complete pagination and no warnings; it excludes code diffs and is not a novelty certificate.
Nearest work: PR1415, unchanged head
0d8d2c3a96346756f02563b401a042ab97198e20, audits HPE operator semantics. PR893, merged head5d8c2e96e0624efbadb182d9ee0a09ee65b70b48, already checks exact command/report coverage for PR236. That general principle is prior work; this contribution isolates and repairs the different HPE driver.Validation
Passed on CPython 3.12.13:
make selected-record-check.Standalone CPython 3.14.6
-I -B -Oreplay matches the committed receipt. The actual Lean 4.21.0 three-probe fixture reproduces exactly. Its deliberate extra axiom is a negative control, not an HPE theorem dependency.The original pinned file already contains the correct 42 queries; these tests do not refute its recorded results. No full HPE/Mathlib rebuild or selected-construction cold replay was run. Full-build compatibility remains an adoption check, and no hosted-CI success is claimed at submission.
Credit
Jeremy Kothe/jeremykothe2 (PR593) supplied the HPE verifier and formal package; DreamingOfClouds supplied PR599 and the Lemma A/G lineage. Their retained notice supplies the construction, realization and Clifford credits. Credit commdata2338/PR1415 and WeberChen8128/PR893 for the complementary audits. No code is imported from those audits.
Prepared for Hanson Sun (@bbMisk) with substantial OpenAI Codex assistance in investigation, code, tests and writing; no subagents or independent human review. Apache-2.0. Emulate was unavailable; prose was checked against sources and executed evidence under the donor's disclosure instruction.
Open next questions
Can this repair preserve the complete HPE build and all 42 actual reports? Apply the proposed patch in an isolated successor and run the original combined verifier with pinned Lean/Mathlib 4.21.0. Full dependency setup and a maintainer decision on successor source pins are required; frozen receipts remain untouched. README/TASK record this boundary.