[roger-sato]

Summary

Answer PR3687 Open question2 with a source-bound, reproducible whole-word cross-check of the RETIMED words in PR3642/PR3653 and the separately serialized DELAYED words in PR3682, against their STAGED control. All six words pass with 632 independent live inputs and 91 independently dirty buffers: complete forward/dual endpoint contracts, all 723-register forward/reverse pull-back coefficients, and the cleanup boundary agree exactly.

This reuses Sagar Thakrar's unchanged PR3687 interpreter, including its explicit Arham Tahir contract/frame-audit lineage. It is not a new interpreter or the first independent retiming proof. PR3642 already proves its local commuting swaps and cleanup reordering; PR3682 independently reconstructs its moves and proves arbitrary-dirty operator equivalence. Sagar also reported an unshipped RETIMED forward/dual/frame scratch check. This package adds committed direct whole-word computation through that evaluator, including the reverse pull-back and DELAYED words, with meaningful failure controls.

• Complete reconstructed JSON equals each published changed word. Both changed profiles are {1:2437,2:724,3:226,4:102,5:20,8:8}, with unchanged widths and rank mass 5135. Each donor 640/656 has delta {1:-1,2:-1,3:+1}; no other register invoice changes.
• All 36 required-outcome controls behave as specified. Omitted, sign-flipped or over-delayed releases fail the actual symbolic endpoint after regenerating cleanup. Old cleanup remains semantically valid but fails literal canonical order. Old release frames preserve the operator but restore the STAGED bill in RETIMED and violate containment in DELAYED. A false declared profile is rejected.
• Six exact compressed word inputs and the original evaluator are retained (~0.5MB total package). Nine unique source mappings, fixed source commits, SHA256 and Git blob IDs are checked before source execution. No foreign Git objects or construction/native code are needed by the direct replay.

Exact heads: PR3605 3efbb5de, PR3642 044d24f1, PR3653 607a09ff, PR3682 431e6c60, PR3687 f798838e. Credit Ross Richardson, roger-sato, Rohan Arun, zxv and Sagar Thakrar respectively, plus Arham Tahir, dcharb78/PR3133, Marius-Paul Dumitrean/PR2979, Jeremy Kothe and the preserved word lineage. Source paths/full commits and the existing coverage are in the package README/SOURCES.

Base 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority revision 6. The independent-moment card is refined to this explicit descendant-word handoff. Fresh open/closed PR and Discussion overlap includes the published local proofs, PR3677 histogram arithmetic and PR3681 inherited full replay; none is described as a first-proof gap.

Test plan

• CPython 3.12.14, macOS 27 arm64: complete result generation passed in 11.58s; direct Git-free copied-package replay passed in 11.30s.
• All 36 semantic/format controls are recomputed with the positive words, not accepted from stored PASS summaries. RESULT.json is compared to freshly computed findings.
• Sixteen bounded authentication checks pass normally/under -O: positive control and rejected missing core, missing word, duplicate/extra mapping, wrong commit, changed core and changed word bytes. The exact retained evaluator and all 9 source mappings were independently authenticated against their original Git/API bytes.
• Committed-byte preflight passes for all 13 files at eb1ecb9e7cc85b4fa8c00eba97dbf8feee9b48b2. The separate full-history check confirms the declared canonical source is in publication ancestry; all foreign word/evaluator bytes are retained and independently authenticated.
• Independent final review passed. The registered normal/optimized pair completed in 22.78s with PASS_REGISTERED_RESEARCH_PACKAGE and outcome=completed, reproducing all six words and 36 controls in each mode. No hosted-CI or complete supplier acceptance is inferred from these local checks.

python3.12 -B scripts/research_package.py --package research/retimed-word-independent-audit
# Direct, also usable outside Git after copying this package:
python3.12 -I -W error -B research/retimed-word-independent-audit/verify.py
python3.12 -I -O -W error -B research/retimed-word-independent-audit/verify.py

This is finite operator/frame evidence only. It supplies no new kappa, selected-frontier change, native Gaussian phase certificate, consumer owner/window proof, root audit, whole-wrapper replay or all-size multiplication theorem. Those obligations remain explicit, including PR3687 OQ1's shared entrance residual. Frozen packages and production helpers are unchanged.

Prepared by roger-sato with substantial OpenAI Codex assistance for investigation, implementation, execution, independent agent review and writing; reviewed 2026-10-19 JST. No human or formal-kernel verification claimed. Apache-2.0.

Hosted follow-up (2026-10-19 JST): the fork Registered research packages run passes at exact head eb1ecb9e7cc85b4fa8c00eba97dbf8feee9b48b2 on Ubuntu 24.04 / CPython 3.12.15. The 265 onboarding-helper tests pass (11 platform skips); selection from canonical base 20ecf289 contains only this package. Both registered normal/optimized commands emit PASS_RETIMED_AND_DELAYED_WHOLE_WORD_AUDIT, the runner reports outcome=completed, and the committed-input immutability check passes. This corroborates the local finite audit; it does not establish native supplier acceptance or any broader scientific claim. Upstream pull-request Actions still require maintainer approval.