Skip to content

Independent verification: point-set replay of κ = 0.001661202643272239 - #777

Merged
Mister-Meeseeks merged 2 commits into
CrocSwap:mainfrom
tekuuu:research/seed60-complex-improvement
Oct 13, 2026
Merged

Mister-Meeseeks merged 2 commits into
CrocSwap:mainfrom
tekuuu:research/seed60-complex-improvement

Conversation

@tekuuu

@tekuuu tekuuu commented Oct 13, 2026

Copy link
Copy Markdown
Contributor

This independently reconstructs and verifies the eight PR699 frame changes giving conditional κ = 0.001661202643272239, up from PR699's 0.001660666354374641. Concurrent PR741 published the same construction first. This is independent verification, not a priority claim or a bound beyond PR741.

The distinct contribution is a standard-library checker that represents subspaces as explicit subsets of the 512 points of GF(2)^9. It imports no constructor or native row-reduction code, verifies every frame inclusion, recounts the complete paid profile (including center descents and final climbs), and confirms unchanged scalar gates and helper widths. Twelve targeted controls reject omitted retimings, a mass-preserving false profile, and scalar/endpoint corruption. Strict concavity proves the direction of the cost change: each edit replaces ranks (2,3) by (1,4).

The focused package includes an independently derived recipe, the unchanged 36,288-owner assignment, frozen native contracts, exact certificate, source pins, search coverage, contributor credits and a fresh incremental replay. It retains the original native, controller, invoice and assembly verifiers. Base: 66849e1475c930fd7732b1a0067ffee7e1e95f60; PR699 dependency: 03e8d7ca494a083e5053cd6b1228f66b98414479; PR635: a6895a69f8ad98ab45777b03e9e6f48888c39f6f. Read-only comparison pins PR741 at e516b1463606840a15f80aac1995d64ce0092e89 and confirms identical input/output word hashes.

After the two dependency archive commands in the README:

python -B research/pr699-frame-retiming/verify.py \
  --main . --pr635 /tmp/frame8-pr635 \
  --dependency /tmp/frame8-pr699/research/gauge44-promo-p6k4sw \
  --work /tmp/frame8-verified

Expected: PASS_PINNED_INPUT_FRAME_RETIMING_COMPLETE_REPLAY 0.001661202643272239.

Validation covers all 728 columns in both orientations, 2,953 phase edges, wrappers, 35 focused frame controls, 95 inherited end controls, 12 independent frame controls, seven independent owner controls, 69 admission controls, 12 bit-supplier controls, two exact moment engines and 47 strict outer inequalities. The selected-record inventory check also passes. The replay starts from a pinned PR699 input freshly reconstructed during this investigation; it does not rebuild the ancestral chain or cited PR589 bit circuit on every invocation. Full receipts are in research/pr699-frame-retiming-validation/.

This remains conditional on the inherited all-size compiler, common-label/mutable-source realization, complete-call hosting, precision/tape, exact-recovery and analytic-transfer interfaces. No full theorem formalization is claimed. upstream/ and the selected frontier are unchanged. Prepared by Tekalign with substantial OpenAI Codex assistance; credits include chafreaky's concurrent PR741/owner construction, jeremykothe2/BRoute's PR699, icekylinx, BlockInvert, Q6Reviewer and the retained upstream lineages.

@Mister-Meeseeks
Mister-Meeseeks merged commit a45436d into CrocSwap:main Oct 13, 2026
Mister-Meeseeks added a commit that referenced this pull request Oct 13, 2026
…ling contributions

Select conditional kappa 0.001661634653669567 from Jeremy Kothe's #799 on Chafik's #741 and the #589/#555 bit supplier. Preserve the previous checkpoint and cumulative attribution. Complete #795, #777, six #681 component reviews, and five exact tooling reviews. Integrate #806 as a scoped hash-substitution diagnostic with an adversarial counterexample to timing-only attribution. Retain source-pinned evidence and separate publication/CI state.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants