Independent verification: JP8 adjoint replay and dirty-centre clear witness

[thakrarsagar]

The selected JP8 scalar endpoint routines initialize 716 persistent input coordinates. This independent reverse-mode checker verifies both original #626 scalar programs on 724 arbitrary inputs, including all eight private centres, and checks all 724 output coordinates. All 492 dirty helpers are restored and private outputs are zero.

It also gives a precise coverage witness: deleting backward event 13113, the initial C0 clear, preserves every ordinary input column and clears all private outputs, yet initial C0=1 corrupts 107 persistent outputs. The package records the entire exact rank-one defect and reproduces acceptance by the pinned predecessor's local() routine with assertions enabled. This is an added scalar obligation, not a claimed defect in the intact construction or acceptance by full production admission.

Status: independent finite verification; no new kappa. Only research/jp8-adjoint-restoration/ is added. No existing construction, selected-frontier record, or upstream/ file changes.

python3 -I -B research/jp8-adjoint-restoration/verify.py

Expected: PASS_JP8_ADJOINT_RESTORATION outputs=724x2 inputs=724 controls=6 kappa=unchanged.

The new implementation imports no production checker or constructor. It pulls each output functional backward through add, scale, copy and clear operations with exact rational arithmetic. Six omission controls produce basis-vector counterexamples, each checked again by forward scalar-value execution. A separate Codex reviewer independently reproduced all 107 defect coefficients directly from the raw events.

Validation passes on macOS arm64 with Python 3.14.8 and 3.11.15, including isolated optimized execution, a minimal relocated checkout in a path with spaces, alternate gzip packaging, source/program/certificate corruption, and make selected-record-check. The README includes a separate reproducible predecessor-comparison command. Full repository/cold supplier replay, Linux CI, Lean, and geometric/physical/analytic re-audits are not claimed; all inherited frame/wrapper, compiler, realization, hosting, cost, precision/tape/recovery, bit-supplier and analytic assumptions remain.

Base: 66849e1475c930fd7732b1a0067ffee7e1e95f60. OS seed 79, independent-verification track, selects JP8 by index 79 % 2 among [ck, jp8]. Source pins bind the selected record and inventory to the #626 SPEC, semantic artifact payloads, program hashes, and reviewed heads. Selected #626: 660de893d27bb7abd1d5b01b1640207528a5c851; #622: f69477f4f6d3c7eb1f740b166881bffed4ffa9f6; #623: e39a615ca601ce5d62abeb24f3fcdb8644ca1005; #543: 2065fe4f72a914d907f1046603f5a36f17018535.

No pinned sprint issue existed, so the claim is local. The full-queue pass covered 553 open PRs through #734, with an immediate pre-submission refresh through #743. Concurrent #734 is credited for arbitrary-private-input coverage in its forward primitive replay; this adds a different execution direction on the unchanged scalar stream and the complete clear-omission defect. #722's coefficient-work sharing and #729's frame-lifetime audit address different obligations. No pending result is imported and no priority claim is made for private-input coverage.

Credits include Jeremy Kothe's #622/#623 construction, Chafik Boukhalfa's #626 package, DreamingOfClouds' scalar/cold-helper audits, icekylinx's compiler/emitter lineage, Dugongue, eumemic's #543 bit supplier, Rohan Arun, and retained predecessors. Prepared for Sagar Thakrar (@thakrarsagar) with substantial OpenAI Codex assistance in research, code, exact checks, review and writing. Apache-2.0; no independent human peer review or formal multiplication proof is claimed.

[Mister-Meeseeks]

100-PR batch review at fc6ebccca3bd61a6eaf6ec3fcbf2e99d2888c35c.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: JP8 adjoint restoration audit adds all eight private-center input columns and exposes an omitted-clear control; retained circuit passes.

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.