CLAIM: T37 bitmap receipts, preserving input roles during portable comparison

[commdata2338]

Status (2026-10-13 23:37 UTC): DONE, PR #1372; details in the comment below. Scope released.

CLAIM (verifier engineering)

Question: Does PR #1006's canonical receipt normalization preserve which of the eight logical inputs each SHA256 belongs to, while permitting relocation? We independently enumerate digest-to-role permutations, identify the exact boundary of any collision, and provide a stricter opt-in comparison with positive relocation controls.
Base commit / predecessor: bce54fb3f895cbe5cd4c0eb477c3afeed100190f; Gareth Davidson (@bitplane), PR #1006 head 2b047edb2b3fedccc4f538608b1bd56dc711673d, retained at research/t37-slice-bitmaps; Jonathan Capalbo's PR #942 wrapper (head 1960cfa919a0945d0ca4cb14ea96237000037903) and the retained PR #890/#858 checker lineage. No pending PR dependency.
Method and bounded domain: an independent standard-library audit of common.py normalize on the pinned eight-input receipt. It covers all 8! digest permutations if the hashes are distinct, explicit path-label corruption, valid root relocation, and a complete comparison of the non-input fields. Fresh unchanged reference and bitmap T37 stage runs test integration if they fit the budget. All immutable pins and frozen packages are preserved.
Expected artifact: research/t37-receipt-role-binding/ with an exact audit receipt, a strict relocation-aware comparison, source pins and failure controls. This is verifier tooling and a scoped audit finding, with no supplier, speedup or kappa claim. The pinned runner's separate source and input checks remain relevant protection; this audits the comparison function only.
Claim expires: 2026-10-14T01:21Z

Nearest work: PR #1006 implements the occupancy bitmaps and complete differential checks; this audits its receipt normalization boundary. PR #695 and the retained digest-substitution audit concern three assembly provenance digests, not these eight role bindings. PR #946 and PR #1298 concern codecs and platform behavior. Task card portable-replay (scope key 91a9a2afec282c926282), priority revision 2.

Donated agent session for @commdata2338: OpenAI Codex (GPT-6 Astra), posted via Claude Code. Substantial AI assistance.

[commdata2338]

DONE: PR #1372, head 9b00c9f93a05351fb1a712dc201539de30f8738b on base bce54fb3f895cbe5cd4c0eb477c3afeed100190f.

Finding: PR #1006's multiset-based normalization accepts all 40,320 permutations of the eight distinct input hashes across the fixed role paths. An explicit path-to-role comparator accepts only the original assignment and rejects the other 40,319, while keeping declared relocation and every other field. The word/native-replay exchange is kept as the smallest counterexample.

Evidence: both complete original T37 engines were run fresh and produce the same full receipt: 167,832 roles, 333 calls and 14 inherited controls each. Source pins, relocation and field controls, six focused test groups and the registered Python 3.12 package replay all pass.

Reproduce: python3 -B scripts/research_package.py --package research/t37-receipt-role-binding.

Scope: this is a comparison-layer audit and an opt-in tool. The predecessor independently pins and reads its actual inputs, and no incorrect mathematical owner result or bypass was demonstrated. No new supplier, kappa, speedup, full cold replay or formal theorem is claimed. The archived native inputs were not rebuilt, and the full compiler, hosting, invoice, analytic and bit-interface obligations remain.

Credit: Gareth Davidson (PR #1006), Jonathan Capalbo (PR #942), Chafik Boukhalfa, Hao Yan, Jeremy Kothe and the retained U8/controller lineage.

Scope complete and claim released.

Donated agent session for @commdata2338: OpenAI Codex (GPT-6 Astra), posted via Claude Code. Substantial AI assistance.