Independent audit of R437 OQ4 transition and staged544 admission replay

[Clear1nsight]

Change and mathematical scope

Adds only research/staged544-r437-admission-audit/: a scoped independent R437 OQ4 template-transition audit and explicitly partial staged544 execution evidence. Original fork review: Clear1nsight#6.

The new standard-library checker imports no predecessor verifier. It authenticates both literal templates, rederives idle windows by last-touch scans, checks 3933 stage/helper slots per replica, chronology and residual suffixes, and reconstructs both complete owner maps including 125856 helper occurrences per head. Exactly 69 semantic leaves change; 32 helper owner descriptors move and the profile delta is {26:32,34:-64,42:32}, preserving rank mass and child count. Every other field, including programs, schedule, cover and stock, is unchanged. Seven mutants are rejected. Modified-program/cover controls pass geometry and fail exhaustive transition, preventing inheritance of old identity coverage to changed programs. Strict concavity proves the improvement direction for 0<c<1 under the unchanged inherited invoice.

This does not discover the slot move, re-prove the full word identity, admit a supplier, establish a new physical hosting theorem or change the selected kappa/frontier.

Task and overlap

Scope: exact-head supplier/admission audit, refined to the OQ4 transition and bounded staged544 evidence; see package TASK.json. Base: 20ecf28; priority revision 6. Scientific head: b122e6c. The upstream comparison has one commit and eleven files. Current AGENTS.md, autoresearch.md and CONTRIBUTING.md were read for submission.

PR3647: old head 36ff441 (covered by PR3700) versus current head a24f9ff. PR3802/T385: de2cd5c. Both templates are retained with SHA256/Git-blob pins. Added value over PR3700 is exhaustive coverage of the actual changed head, plus explicit execution-stage boundaries. Relevant PR overlap was checked in the original investigation; Discussion coverage/posting was unavailable then. No exhaustive novelty claim or new reservation is made.

Validation

python3.12 -B scripts/research_package.py --package research/staged544-r437-admission-audit
python3.12 -O -B research/staged544-r437-admission-audit/verify.py
python3.12 -B scripts/check_research_commit.py --package research/staged544-r437-admission-audit --check-source-history

Recorded publication/bounded T385 run: https://github.com/Clear1nsight/integer-mult-bounds/actions/runs/39188328699. Metadata, registered replay, optimised receipt equality, history/publication bytes, Git-free replay and CLI altered-input/result, missing-input and symlink controls passed at the scientific head. Original execution: https://github.com/Clear1nsight/integer-mult-bounds/actions/runs/39186343975.

R437's existing cold verifier passed all 13 commands under hosted Python 3.12.15 and local 3.13.5; c=1736646042038446578551/10^24 and next-grid rejection were cross-checked with the unmodified PR3700 rational engine, and the old profile at new c was rejected. These reused-code runs are distinct from the new transition implementation.

The full T385 wrapper reached its 780-second limit in discovery/audit (exit 124). Owners, parent-lift, rebind, discovery bind and emit passed. Status remained PREPARED_REPLICA_SAME_WORD_DERIVATIVE_NOT_ADMITTED; no fresh full discovery, outer stages or release pass is claimed. A later bounded Python 3.12.15 verifier passed the exact aggregate, 1413720 literal roles, three owner controls, regenerated outer ledgers and conditional-gain checks. This authenticates stored evidence; it does not convert the timeout into a complete replay. These are prior recorded executions; no new scientific execution is claimed by this browser publication.

Credit

Credit ai-nanachi/PR3647, Sagar Thakrar/PR3700 (old-head auditor, OQ4 placement and reused pricing), Ali Romman/PR1043, dataisfire/PR3734, hexapode/PR3802, ChaloupeMusic/PR3815 and the retained native/owner/outer lineage. New checker, controls, execution and writing by Chris R / Clear1nsight with substantial OpenAI assistance. No independent human or formal-kernel review. Apache-2.0 and source notices retained.

Open next questions

See package README.md and TASK.json. Complete a longer fresh T385 replay through discovery, both outer stages and release, retaining exact source and explicit reuse coverage. All-size compiler, mutable-SOURCE, arbitrary capture-stream, whole-call hosting, precision/tape/recovery, analytic and BIT interfaces remain conditional. These are remaining obligations, not admission claims.

Workflow feedback

This separate upstream PR transfers fork #6 after the integration's prior HTTP 403 prevented upstream creation. The fork record remains the historical execution handoff.

[Clear1nsight]

Follow-up completed in #3936 (125f5af). Fresh pinned T385 wrapper now passes all seven stages through release: PASS_T385_COMPLETE_CONDITIONAL_RELEASE, exit 0; 2946.086 s recorded supervisor interval, 2923.426 s scientific stage sum, explicit native/ancestry reuse and runtime adaptations retained. This supersedes only the earlier execution gap.

The new independent R437 checker covers exact physical pullbacks on all 659 arbitrary input functions in both programs, seven semantic controls, a dirty-input-340 separation from the weaker named-dirt contract, and complete finite COPY/disposal accounting under the inherited tariff. Existing source-preserving caller code is traced in CALLER-BOUNDARY.md; its uniform cost/scratch, routing, volume, precision/recovery and theorem premises remain conditional. No selected-frontier change. Normal/optimized/Git-free checks and published-byte/source-history checks pass; CI queued at publication.

Completion handoff posted under Discussion #1166, comment 19030047. Credits and unchanged predecessor notices are retained; substantial OpenAI Codex assistance disclosed.