Contract revision response

Preserved research record. This public derivative removes local-only path links; scientific wording and original status claims remain historical. Internal AI reviews are not external peer review. Original SHA-256: 72fd5305deee6e6a3192758972c4755feb7b6920207c2bf7d7447bbc6b0a987d.

Equations in these preserved Markdown records retain their TeX notation where present; this is not a newly typeset proof edition.

Response to the attached publication-readiness review

Revision: 26 September 2026. Disposition: contract/specification corrections completed; synthetic pair-token primary test failed; natural admission/correctness and bounded secondary timing completed; full systems and submission gates remain open.

This response now addresses both the pasted summary and the subsequently supplied full audit/checker bundle. The full report was read, its 21 manifest entries were verified, and its unmodified reference checker was rerun with results identical to the supplied record. FULL_AUDIT_RECONCILIATION.md maps every P1/P2/P3 finding to current evidence and remaining work. The September 22 manuscript, prior audits, and first revision delivery ZIP are preserved. The compiler source remains unmodified.

Review concern	Change / evidence	Status
Dense output incorrectly described as unfixed-axis output	Sections 5.1–5.4 now retain every ambient row/column axis, distinguish syntactic/essential/ambient counts, give exact shape and the 4-by-2 fixed-Z example, and charge NumPy Boolean bytes. Tests cover both fixed values, all three strategies, irrelevant axes, ownership and budget refusal.	Resolved in draft and executable regressions
Valid derivations confused with selected strategy	Section 4 treats inference as an admissibility relation. Section 5.2 gives ordered case selection, S/T/H outcome mapping, and the legacy `structural` alias. `X AND Y` can derive S and T; pure compilation of `X OR (Y AND Y)` still fails. The theorem's fallback clause now explicitly concerns selected failure.	Resolved
Support ambiguous	Explicit recursive `Vars(e)` minus `dom(F)` definition, with cancellation examples. No essential-dependence minimization is assumed.	Resolved
Child folds assumed to survive root failure	Section 5.2 states original-AST fallback and distinguishes attempted counters from committed results. A spy regression verifies the identical original AST is passed to the real builder after child success. Root retabulation can also supersede earlier work.	Resolved; no different persistent-rewrite algorithm introduced
Token failure, invalid input and actual fallback conflated	Duplicate/overlapping layouts are errors; token `None` with `ordinary_fallback` is a diagnostic; only the dense wrapper executes fallback. Budget checking occurs before token compilation.	Resolved
Novelty too close to classical operations	Abstract and contribution language narrowed; Proposition 3.3 attribution, kitty operations, mockturtle's ordered leaf interface, and AIG candidate/gain/commit behavior compared explicitly. No claim of novel truth-table algebra, unprecedented operand metadata, or measured engineering advantage.	Positioning corrected; research significance remains for reviewers to assess
No current implementation artifact/check chain	Recovered 63 source/import/test files at the pinned commit, checked Git blobs and SHA-256, added a portable runner, 33 contract cases and current manifest. 90 upstream tests plus the 33 new cases pass.	Current contract chain closed; historical 116-test identity remains unestablished
Historical evidence merely reported	Recovered 95 byte-identical S1/S2/P14 archive members. Recounted S1/S2 observations and all P14 primary point estimates; reproduced all 17 P14 bootstrap interval sets from 51,102 raw rows. Eight P14 freeze hashes match.	Analysis closure improved; historical imported-runtime identity is not certified by its freeze
Missing/obsolete package-relative checker	Current runner uses only packaged dependencies and data. Historical broken checker is retained as history and clearly not the current entry point.	Resolved for this revision
Confirmatory pair study missing	Preserved protocol v3. A disclosed synthetic-only protocol v4 retained its primary cell and thresholds, froze 64 held-out formulas, and executed 30 paired repetitions per formula. Direct packed/CM median time ratio was 0.2479 (95% interval 0.2462–0.2495), so primary utility failed.	Synthetic cell answered negatively under a post-import cache-reset regime; full v3 circuit scope remains open
Natural applicability unknown	Frozen admission/correctness screened 2,861 original EPFL outputs and admitted 71 cones from ten designs; hybrid roots were four pure, one retabulated, 66 fallback. A separate frozen descriptive timing run completed 70 paired full-bitset cases and retained one calibration failure.	Bounded selected-cohort evidence; broader prevalence and matched systems performance remain open
Other useful editorial/evidence corrections	Removed detailed later-project results from abstract; kept P14 table with its subsection; corrected 98.75% marginal versus 95% familywise interpretation after inspecting frozen analyzer quantiles; restored P14 environment/amendments; corrected AIG page range and Lee first author.	Resolved in revision

Verification actually performed

All 123 selected tests pass under Python 3.10.11 / NumPy 2.2.6 / pytest 9.0.2, including 327,680 signed-fusion assignment comparisons across all five implemented outer operations.
63 pinned source files have verified Git blob identities and current SHA-256 checks. They are not edited to make the tests pass.
Recovered S1/S2: 10,500 rows, 10,389 all-arm completions, zero differences in the seven stated CM/generic symbolic metrics, and the reported failure patterns and medians confirmed.
Recovered P14: 51,102 timing rows; four primary case-balanced ratios and all 17 archived interval sets reproduced with 10,000-draw bootstrap replay; eight frozen source/holdout hashes verified. This did not rerun historical timing workers.
Frozen v4 primary cell: 64 held-out formulas, 3,840 timing rows, 128 memory rows, zero correctness disagreements or recorded failures. The independent recount matches the frozen analyzer. Median packed/CM ratio 0.2479, 95% interval [0.2462, 0.2495], so primary success is false.
Frozen natural secondary run: 70/71 cases completed 15 paired Python full-bitset repetitions per arm; one adder case failed the fixed calibration floor. Direct packed evaluation was faster on all 70 complete cases, with median case-level direct/CM ratio 0.0376. All 71 cases completed 15 ABC conversion/truth repetitions. An independent verifier checks counts, hashes, truth digests and median arithmetic.
The PDF build and visual QA records are packaged separately. A successful build is not a mathematical proof or a publication-readiness certificate.

One initial new checker attempted to pass an arbitrary outer truth-table integer to the implementation's string-only five-operation cm_compose API. The checker was corrected to test the five implemented operations, giving the explicitly reported 327,680 comparisons. No implementation behavior was changed. The first LaTeX build encountered a sandbox denial for MiKTeX's user cache; the build was subsequently run with the required tool access.

Remaining submission work, in priority order

Complete task-matched ABC/AIG, ROBDD and generic-folding comparisons for natural inputs. Map sharing, reuse, output and process-start regimes without replacing the failed v4 primary cell.
Establish which imported compiler/runtime bytes produced historical P14 timings, then prepare a public versioned archive with corpus provenance and a stable identifier. Bootstrap analysis replay is complete; new timing workers are not needed for that analytical step.
Reassess venue and contribution significance using the resulting evidence. The contract suite is a useful methods artifact but does not alone prove journal-level novelty or performance.

These are separate research/release requirements. The authorized draft revision, full-audit reconciliation, source recovery, regression checks, historical bootstrap replay, synthetic v4 timing campaign, natural correctness gate, and bounded natural v5 secondary timing are complete. The audit's original readiness score is a historical assessment of the September 22 draft; it has not been silently recalculated or presented as a new independent verdict on this revision. No production compiler code change, external publication, commit, or push was performed.

Later natural-circuit correctness and timing addendum, 26 September 2026

The separate natural extension screened 2,861 original EPFL primary outputs under rules fixed before CM compilation, admitting 71 from ten designs. The pinned ABC executable converted all 71 to truth-equivalent AIGs. A frozen correctness run passed all 71 cases against the original BLIF with direct packed AST, ordinary CM IR, ABC AIG, and every returned pair token. Under hybrid, four selected roots were pure structural, one was fully retabulated, and 66 fell back. The subsequent frozen descriptive secondary timing run completed paired Python producer timing for 70 cases, with one retained calibration failure. Its case-median direct/CM ratio is 0.0376, and direct packed evaluation was faster on all 70 complete cases. ABC conversion and exhaustive AIG truth checking completed 15 repetitions on all 71 cases, but ABC subprocess wall time is not a matched ratio against resident Python producer times. This does not alter the failed v4 primary result or resolve P14 historical runtime identity. ROBDD/generic comparisons, sharing/reuse/output regimes, independent-family prevalence, and a full matched systems campaign remain open.