[kumagi]

PR #5521 explicitly left PR #5369's reversible CNF reduction outside its independent audit. This package reviews PROOF.md §6 and the actual reversible_cnf.py at source head 18354f41de4dc7bfda4692e814cf1d91a6d619ef. No defect was found under the stated unbounded five-role commuting-sign grammar. The independent proof covers dirty controlled swaps, chronological Boolean compilation, empty formulas, instruction bounds and exact rational bit complexity.

The audit makes the encoding qualification explicit: ≤1200·max(1,m)² instructions is not a quadratic byte bound. The source emits O(n³)-bit dense generator tables; renumbering used variables to n≤max(1,3m) first makes the whole reduction polynomial. The source compiler does not automatically perform that renaming. Inequality certificates need only referenced generator signs.

A separate Fraction interpreter checks actual source-emitted instructions and all five dirty input columns. Evidence includes all 14,400 permutation branch pairs / 28,800 endpoints, all 24 five-cycle targets, 1,807 CNFs / 14,456 truth assignments, 820 complete tensor blocks at widths 1–3, a 201-bit variable-name compaction fixture, and eight semantic mutations each exposing a specific wrong matrix entry. The source evaluator/verifier is not used as the oracle.

Validation passes: Python 3.12 registered and isolated optimized replay; source pins; committed-byte/source-history check; deterministic 16-file export/integrity check; fresh Git-free optimized replay with the same receipt; altered-source and altered-receipt rejection. See research/reversible-cnf-independent-audit/VERIFICATION.md. Current canonical main remains the pinned base df0137efd35e11b4445734bf612bdd4c7ebc330c.

Known shared CI limitation: the unchanged test_research_current_status.py runs six tests and fails the two review/source_manifest navigation subcases because docs/ci-verification.md lacks selected links. PR #5703 already supplies that repair. This PR changes only its new package and does not duplicate the shared fix. Passing local scientific evidence is separate from hosted planning/replay acceptance.

Scope excludes the general native compiler/solver, the four-width cutoff theorem, physical query costs, source event8814, supplier admission and the full multiplication theorem. No new hardness theorem or κ; the selected P104 result and all five inherited premises remain unchanged. This is agent-assisted independent review, not human or proof-kernel acceptance.

Canonical claim and handoff: #5354 (comment)
Original explicit review request: #5354 (comment)

Ali Romman retains the native reduction and compiler, Barrington the classical width-five method, and Mihir Srivastava the preceding audit and handoff. All source predecessor attribution is retained. New proof, code and review prepared for Hiroki KUMAZAKI with substantial OpenAI Codex assistance and delegated proof/interpreter review. Apache-2.0.