[mihir-s-05]

The native-word handoff from #4478 still lacked an independent chronological frame check for #4470. This adds research/pr4470-frame-audit/: a separate GF(2) elimination and per-field demand-path checker, compared against the exact source constructor's emitted programs in both orientations.

Result: all 499 edited locations bind to their predecessor (38 entrances, 47 exits, 414 contacts). Both orientations reproduce 3,363 → 3,310 paid jobs, with paid and helper rank masses each increasing by 37. All 408 helpers have strictly nested productive endpoints; all eight rank-eight center jobs remain charged. The checker covers 17,453 scalar events and 240 warp events per orientation, their coefficients/labels, chronological source states, final frames and retained owned operations.

The source constructor only emits the object under test. Its proof engines are not called by the independent checker. Nine byte-exact source snapshots make the replay self-contained; most added lines are these unchanged, attributed inputs. Start with the package README and verify.py; the complete deterministic result is in RESULT.json.

Validation on CPython 3.12.0 / Windows 11:

• Registered package replay passes, including eleven semantic rejection controls and both predecessor/changed positive cases.
• Isolated optimized replay (python -I -O -B) passes.
• Committed-byte preflight passes for all 20 package/input files.
• A committed export passes outside any Git worktree in 9.5 seconds; corrupting its source PATCH.json is rejected before constructor execution.

python -B scripts/research_package.py --package research/pr4470-frame-audit

Scope: independently checked frames, retained event word and local billing. The frame-coordinate implication still assumes the physical transitions implement the stated operators. Exact Clifford phases, helper normalizers, full arbitrary-column identity, the p16 controller, precision/stock and supplier admission remain separate. Linux execution and hosted CI were not run locally. No new κ; selected κ remains 23/10000.

Base: canonical main 2a13eb1827784206202db90e5dc8327cf8a250db, priority revision 7. Audited #4470: b9a13c1673e0e9697de410de16e8bbf3207a7363; predecessor #4447: 322e0ec764aa7ef7801383feb1f656e460dd8bed. Main and relevant heads were unchanged at submission refresh. Live overlap included fully paginated open PRs/Discussions and the complete local top-level note inventory. #4478's current verifier already covers uniform gain, so that arithmetic was not repeated.

Canonical claim, findings and next tasks. Validation record.

Credit Cinna Davis / 1337isnot1337's construction and native implementation; zxv, Jacob Alan Sussman and retained native/geometry contributors; Ryanson Jonathan / pinterbanget's independent arithmetic audit and handoff. Prepared by mihir-s-05 with substantial OpenAI Codex assistance. Apache-2.0.