Independent verification (no κ): PR3633's T488 journal, fixed-inventory ceiling and endpoint reproduce; inherited owner-audit controls cannot fail

[thakrarsagar]

Summary

This is an independent verification of hexapode's open PR #3633 (T488, head 72c268f1). It claims no new κ.

PR #3633 combines three chronologies with Ross Richardson's staged 644/660 supplier (#3605): T37 from Ross Richardson, T8 from hexapode, and T16 from Nick Petalas via hexapode. It claims κ = 0.001730096363528623534230 − 10⁻⁴⁶ − 10⁻⁷⁰, and that its allocation attains the fixed-inventory packing ceiling that #3630 bounds.

The new package research/t488-independent-audit/ is standard-library code that never imports or runs PR3633, PR3630 or PR3605. In the checked scope it finds no correctness defect in PR3633's finite claims:

• Literal journal. All 1,791,936 owner roles pass a separate audit:

  • windows derived two ways from the frozen template;
  • PR3630's active-local-pair exclusion, which the windows imply;
  • same-call physical families, capacities and chronological prefixes;
  • the complete invoice over 983,808 bins, which equals the claimed profile.

  The journal ships compactly (0.77 MB) and is bound to PR3633's payload pin by byte-exact re-serialisation.

• Fixed-inventory ceiling. An own LP (2,551 patterns, exact simplex, rigorous dual) shows that within the fixed model no allocation for any replica count has a larger endpoint. PR3633's profile is exactly 488 × the unique LP minimum, so it attains that ceiling. The LP also cross-checks PR3633's CEILING.json and PR3630's dual and digest.

• Supplier words. Both supplier words are exact, forward and dual, in this contributor's word lineage. Their widths equal PR3633's providers, and the change is localised against the PR3133 parent words. This account's parallel session independently audits the same words in Independent verification: PR3605's staged 644/660 words pass an exact audit against the PR3133 control (no κ claim) #3687, and the two agree.

• Endpoint. The endpoint and κ are reproduced in two rational enclosures, with the exact gains over Combine T37 and T16 chronology: conditional κ > 0.00173007186470 #3626, Combine T37 and T8 chronology: conditional κ > 0.00173004755209 #3624, Bound staged-supplier packing for every replica count #3630 and Reduce helpers 644/660 from width 8 to 7; conditional κ > 0.00173 #3605.

• Downstream. rohanarun's #3653 puts roger-sato's #3642 retimed words, whose widths are unchanged, on this journal. Its endpoint reproduces, and the dual re-certified at that endpoint puts it at the retimed ceiling.

Two process findings, both non-blocking:

1. Tautological controls. The five reject() controls kept in PR3605's adapted owner auditor cannot fail. The same holds in all 15 located copies, across Reduce helpers 644/660 from width 8 to 7; conditional κ > 0.00173 #3605, Combine T37 and T8 chronology: conditional κ > 0.00173004755209 #3624, Combine T37 and T16 chronology: conditional κ > 0.00173007186470 #3626, Bound staged-supplier packing for every replica count #3630, Combine T37, T8 and T16: conditional κ > 0.00173009636352 #3633, Match the staged-width conditional κ frontier with a smaller T360 allocation #3638, Improve conditional T530 κ through two commuting release retimings #3642 and Compose retimed T488 suppliers: conditional kappa 0.00173041889976 #3653. The gates are sound and are re-checked here with genuine controls.
2. Platform-dependent native pin. The native_canonical pin hashes the gzip container, and a fresh Linux rebuild differs only in the header OS byte.

Verification

• python3 -I -B research/t488-independent-audit/verify.py --fetch: PASS. RESULT.json was reproduced under both CPython 3.14.4 and 3.12.15, with 212/212 negative controls rejected by their named rules, in about 3 min.
• python3.12 scripts/research_package.py --package research/t488-independent-audit: PASS_REGISTERED_RESEARCH_PACKAGE.
• python3 scripts/research_package.py --package research/t488-independent-audit --check-only: PASS_PACKAGE_METADATA_AND_PINS.
• verify.py --fetch --mutations --jobs 2: PASS_MUTATIONS. All 45/45 checker mutations were killed after a passing unmutated baseline (CPython 3.14.4, on the final code bytes).
• Local-only evidence. PR3633's own fresh end-to-end replay.py was attempted four times. Fresh native, owners, parent-lift, rebind and bind passed; discovery emit was killed by the OOM killer above 5.1 GB on this 7.9 GB machine.

Scope and limits

This PR does not check the native lowering beyond the free-operator replay, the release stages (precision, prefix receipt, outer transfer), the BIT side (only compared), or the inherited all-size hypotheses. It proposes no selected-frontier change.

Coordination

• Claim, answering Clear1nsight's handoff.
• zxv's concurrent #3638 (11:58 UTC, before the claim) reaches the same endpoint at T360; it is credited and was not audited here.

Sagar Thakrar (thakrarsagar), with substantial AI assistance (Claude Code; subagents built two scopes and ran two adversarial review rounds). No independent human review is claimed.

🤖 Generated with Claude Code