Skip to content

Copied centers and reversed corners for conditional κ = 3.850771033e-5 > 2^-15 - #37

Merged
Mister-Meeseeks merged 7 commits into
CrocSwap:mainfrom
rohanarun:research/copied-reversed-corners
Oct 8, 2026
Merged

Mister-Meeseeks merged 7 commits into
CrocSwap:mainfrom
rohanarun:research/copied-reversed-corners

Conversation

@rohanarun

@rohanarun rohanarun commented Oct 8, 2026 •

Copy link
Copy Markdown
Contributor

Under the retained analytic and fixed finite-alphabet multitape hypotheses,

T(n) = O(n (log n)^(1 − κ)), with κ = 3850771033/10^14 = 3.850771033 × 10⁻⁵ > 2⁻¹⁵.

This composes icekylinx's copied-center schedule (#36, 11817cc) with James Chang's reversed boundary family (#34, 7fecbe3) at dimensions (23,25) and the attributed RaD balanced assembly. It improves #36's reported κ by 0.132122792%; this compares conditional exponent savings, not measured runtime.

Change and proof

The data profile becomes 9 singletons + [21,17,481], replacing 11 singletons + [21,15,481]. Every copied-center transform and all 4,073,300 paid rank-one endpoint corrections remain. Original producer choices, role volume, total rank and maximum child are preserved. The new rank-one cleanup complements are included in the same common-basis argument.

The exact bit saving is 962729831/25000000000000. The unchanged #36 complex saving is 717/10^7. The semantic envelope explicitly retains the scalar gate charge G; C1=1 and product row degree 2000. All 47 strict constraints and seven margins pass, with a positive final absorption gap.

Checked before filing

  • All 25 controlled permutation completions, both restriction trees, actual local/auxiliary contractions, all 47 rank cuts, and three exact rational pivot witnesses.
  • Complete physical accounting and an independent class-by-class reconstruction; the old profile fails at the new saving by a rigorous lower moment bound.
  • Copied-center schedules in both orientations on arbitrary dirty inputs, including phase-copy directions and paid endpoint correction; scoped independent agent reviews are included.
  • 12 focused tests pass, including corrupted source/rank-cut controls, rank-preserving child-list corruption, old precision envelope, invalid rows, and next-grid lower exclusion.
  • The focused patch reconstructs 19 files byte-for-byte from pinned Copied retained centers and two-stage assembly for conditional kappa = 3.84569e-5 #36.

Full rerun passed: make verify completed successfully on research commit cb86e50e9a07685068874d8e4174b2e6c209b95c: 182 tests, fresh selected h28/d19 complex producer, inherited h25/h23/h57 scalar and positive-label checks, historical complex h28 reconstruction, and all 18 historical patch checks. The new 12 focused tests pass as well. The validation receipt records completion and log SHA-256 020a0bb269ff326b34056e893484eb408ce0c06c3cae9ced920f64ba1f491f2e. Only an inherited JSON key-order difference was generated; exact parsed equality was verified and the pinned historical file restored. The validation follow-up changes documentation and the receipt, not the scientific sources. The refreshed focused patch reconstructs 20 files. Ready for review; no formal verification or external expert acceptance is claimed.

Start with the proof and reproduction guide. Run make copied-reversed-check, make copied-reversed-producer, and make verify. The inherited analytic, tape, routing, recovery, constructive setup and eventual-threshold hypotheses remain explicit.

Attribution

Rohan Arun, with substantial OpenAI Codex assistance, supplies this composition and checks. Credit icekylinx (#36 copied centers/complex graph and earlier producers), James Chang (#34 reversed family), RaD/hipotures (balanced routing/semantic/bulk transfer), Dominik Scholz (#33), Rohan Arun/Codex (#31), Zhihao Chen (#29/#21/#23), Aurel Prosz/Paureel, Swapnil Jain, eumemic, Douglas Colkitt, OpenAI, Harvey–van der Hoeven and all retained source notices. The inherited techniques are not claimed as new here. Apache-2.0; original licenses and AI-assistance disclosures remain.

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