Skip to content

Validate cover totals at consumption without breaking pinned RN sources - #45

Merged
d6g8k5htny-coder merged 1 commit into
chatgpt/drive-github-hardening-20260919from
chatgpt/cover-accounting-audit-20260924
Sep 24, 2026
Merged

d6g8k5htny-coder merged 1 commit into
chatgpt/drive-github-hardening-20260919from
chatgpt/cover-accounting-audit-20260924

Conversation

@d6g8k5htny-coder

Copy link
Copy Markdown
Owner

Concrete accounting repair, three files

Independent follow-through on Claude's #38, coordinated in comment5804954866. At hardening6f0f061 I reproduced two live accounting omissions: (1) a stored OUTSIDE residual [1000000,1000000] was skipped while total() returned a certified [1,1]; (2) an unrecognized mutable terminal disposition was omitted from both contributing/rejected sets while the exact spatial partition still passed. These are internal consistency defects demonstrated on synthetic ledgers, not a proved error in a historical RN field bound.

New research.cover.audit.checked_total validates record identities/order, dispositions, rejection kinds/reasons/areas, exact interval fields and canonical OUTSIDE residuals immediately before consuming a total. It independently adds Fraction endpoints and compares the reported interval, areas and coverage/certification flags. Existing partition/refinement checks and PENDING refusals remain; well-formed partial and noncertifying results remain partial/noncertifying. tools/rn_side24_spatial_check.py now uses the checked boundary. A real call-site sentinel test prevents accidentally bypassing the adapter.

Why an additive boundary rather than another pinned-source edit

inner_wedge_20260920_v1.json binds research/cover/ledger.py twice at37974B/SHA25658ebe18fbc3fc44e1ba9566e19494fbf906b8579e5dca96b414fb0f309d866c9. Editing that historical module to transport #38 verbatim would break existing replay custody, just like the ladder collision. This change leaves ledger.py, driver.py, regions.py, init.py, all candidates, source pins, registers and scientific statuses byte-identical. No silent re-pin or historical result rewrite. The old raw API still exists: callers must use the new checked boundary for this protection. This PR wires the current RN rectangle replay, not every historical consumer.

Requiring OUTSIDE residual None/[0,0] is a canonical interface choice. A wider interval containing zero can be conservative and is not a mathematical contradiction; it is rejected here so the publication route never silently discards a stored interval. Supplied geometry, integrand bounds and certifying-path assertions still require their own justification. This adapter is not a concurrency lock, authentication system, proof assistant or universal ledger audit.

Executed local validation

Python3.13.5, pytest9.0.2, no optional pytest plugins: 100 passed (59 retained cover/upper-budget cases +41 new cases), zero skips/errors,12.14s. Actual Cartesian and polar reference-driver totals remain unchanged; signed residual sums, incomplete regions and noncertifying results are retained. Normal and optimized Python both refuse the adversarial OUTSIDE record.

Four deliberate mutants were run in disposable copies and each failed selected assertions: remove the OUTSIDE guard; omit unknown-disposition validation; skip independent interval comparison; bypass the adapter in the actual spatial consumer. No pinned source was mutated. The old snapshot archives were hash/Git-blob verified before reuse; live API confirmed matching cover modules and the current consumer preimage8c6bb730. All3 uploaded Git blobs match the locally tested files exactly.

Candidate f8bac0e, tree0e6453024f6b99cc3dd1632a0eba4bff26140e08, base6f0f061588bfd10d1d7b9228d858e1df5d0c3290. Full current-tree RN replay and repository CI run separately; they are not claimed passed at opening. No workflow, dependency or environment permission is changed. Existing automatic pytest discovery covers the new tests.

Authored by OpenAI/ChatGPT. This advances executable accounting, not OBL-H5-JETMOD or RN-UNIF closure. No renewed Dylan approval requested.

Add record and independent sum validation; route the RN local spatial checker through it. Reproduced nonzero OUTSIDE omission and unknown-disposition omission. 100 focused local tests and four disposable mutant checks passed. Full repository CI remains a separate execution.

Authored-by: OpenAI / ChatGPT
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

Copy link
Copy Markdown
Owner Author

Execution evidence is now in Drive, not just the PR description. Package: https://drive.google.com/file/d/1rvj1qvvnWmd2n6bnhPeQmcCFeiesRbwu/view — 25,369 bytes, SHA256 cfdb5ae2a7b9ccbb8ef11feab26d6714c89b88fc9cb3fa630d37f7d666221aae. Readable scope/reproduction note: https://drive.google.com/file/d/1WwhPX2CREmt2sP0076CmczkqzWCnEsJx/view — 3,712 bytes, SHA256 c817991075fd8637719e797352846d0dd6ce90019628820a622854f88c0456eb.

Both were downloaded again and matched exactly. All 15 manifest-listed payloads verified. In a disposable copy, the downloaded patch passed git apply --check, applied to the confirmed original consumer, reproduced all three submitted Git blobs exactly, and passed 100 tests again (12.71s). Original 12.14s JUnit and all four caught-mutant logs remain inside the immutable package; the second run is a separate readback check, not added to the count as distinct test coverage.

Actual PR Actions35936758657 has passed the RN local rectangle replay; full run remains separate. No secret URLs, private manuscripts, new environment permission, source-pin changes or workflow suppression were needed. The automatic code-review bot's quota refusal is recorded as unavailable review, not approval.

@d6g8k5htny-coder
d6g8k5htny-coder merged commit 74c082e into chatgpt/drive-github-hardening-20260919 Sep 24, 2026
4 checks passed

Copy link
Copy Markdown
Owner Author

Integrated; validation scope preserved

PR45 is merged into the active hardening branch at 74c082e. Its full PR workflow 35936758657 and PR withdrawal controls completed SUCCESS at candidate f8bac0e/base6f0f061. The duplicate push workflow was still running when inspected; no separate push PASS is asserted.

Readback exposed a concurrent base advance: the actual merge first parent is 3d47d1b, not6f0f061. I compared them: one commit, additions only in four RN/research-map Markdown documents, no overlap with this three-file code patch. Git preserved that work. The actual merged tree is 348650239ee4d7d80a5bc96dd0d242f2c596c11e, distinct from the tested candidate tree; therefore the earlier CI is NOT labeled as a fresh execution of this final combined tree. Post-merge verification remains a separate observation.

The accounting adapter and its current RN consumer are now shared code, not merely a draft. Historical ledger bytes/candidate pins and the separate R3 migration are not changed. Drive evidence and downloaded-patch replay remain as recorded above; their earlier pending-CI wording is a dated observation, not something to rewrite retroactively.

Copy link
Copy Markdown
Owner Author

Post-merge execution gap closed by actual existing CI: push run35937914517 / verify job107439061362 completed SUCCESS. Its checkout log names the actual merged commit74c082e799d8d338da9adb24a968c1231842851a (including Claude's four documentation changes); pytest reports3169 passed,2 skipped in781.69s. This is a distinct run of the merged tree, not a relabel of the earlier PR run or PR21's3486-test branch.

Continuation scope: I found the second live ledger consumer in tools/rn_inner_wedge_check.py:assemble_cover still calls raw ledger.total(). I am wiring that consumer through the existing checked_total and adding actual call-site corruption/bypass controls. Frozen ledger/driver/region modules, candidate identities, scientific statements and source selectors will stay unchanged. This is bounded engineering follow-through; Claude retains PR21 R3/source-route migration and H3 filing. — OpenAI / ChatGPT

d6g8k5htny-coder pushed a commit that referenced this pull request Oct 19, 2026
d6g8k5htny-coder pushed a commit that referenced this pull request Oct 19, 2026
…#435 and query-#45 records

Adds flb_original_check_20261019/: the extraction script and log that
reproduce ORIGINAL's identity from Math-#193/6242350213, the item-by-item
comparison record, and exact sympy checks of the shared tensor and
regression identities in d = 3 and 4. The DELTA records the main#435
R92-B review, the query-#45 correction and stand-down, and the FLB
pickup. Scientific effect NONE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015wNj8LPTKXsaT68G3DgPPh
d6g8k5htny-coder added a commit that referenced this pull request Oct 19, 2026
Session-initiated retention of what this session posted after increment 5:
the handoff refresh (query-#45 reconciliation, the HC wording correction,
the C293 clarification, the board refresh) and the METHOD V2 section 2
source-fidelity input (posts, readbacks, the pre-exposure note whose
SHA-256 the result cites, check scripts and outputs, and an extraction
script for the ORIGINAL bytes). Not for merge; scientific effect NONE;
credit 0. SHA-256 sums included.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017Mi3hxjaxV45x6zo6o1ee3
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.

1 participant