Skip to content

Certify the exact rational split optimum for fixed P104 flags - #7765

Open
aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-078
Open

aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-078

Conversation

@aliromman

@aliromman aliromman commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

Exact finite optimum: rational split gathers reduce the junk count from 4,814 to 4,745 on the fixed eight flags of PR7748. The full schedule improves 44 columns while preserving 104 diagonal entries equal to 2 and all 5,720 required odd off-diagonal zeros. Column 4 improves from the published split value 44 to its exact minimum 43. No new kappa follows.

Read the proof, 104 column certificates and schedules, checker, source pins and credits. This resolves the fixed-flag optimum left open by PR #7748. The flags, chain pairs and one-read rule are fixed; arbitrary rational legal gather pieces are allowed, with total one on each active chain.

The prefix coefficients describe all rational schedules, without a finite coefficient search. Exactly 4,715 junk entries are constant. Twenty-three integer contradiction certificates force 30 additional entries. All 1,400 equation subsets agree under direct rational elimination and the separate contradiction test. A matching full schedule proves the lower bound is sharp. Exhausting all 1,242 single-gather column schedules gives the comparison minimum 4,814.

The witness uses 490 nonzero pieces across 416 active column-chain pairs: 74 extra pieces. Its piece count is not claimed optimal, and physical movement and restoration are unpriced. The fixed flags remain worse than the 4,216-entry witness reported in PR #7689; that global witness was not replayed here. This result is not a bound over all flags or a paid supplier improvement.

Python 3.12, standard library only:

python3.12 -B scripts/research_package.py --package research/p104-fixed-flag-split-optimum

Both registered commands print:

PASS_P104_FIXED_FLAG_SPLIT_OPTIMUM
104 columns: exact rational split optimum 4745; single-gather optimum 4814
4715 forced cells and 23 contradiction certificates force 30 more; 1400 subsets agree
490 gather pieces; 104 diagonals and 5720 odd zeros; source column 4 improves 44 to 43

All 10,816 affine forms and the complete schedule pass the source formula, affine evaluation and direct piece accumulation. Four rational prefix families give 416 further column checks. Controls reject a false contradiction, a lost optimum, a late gather and a wrong total; omitting a needed contradiction weakens the bound. The unchanged source replay, normal/optimized registered replay, syntax/indentation, committed bytes/source history/modes, export and sealed read-only Git-free serial replay pass. Hosted CI did not certify this package. The plan job failed in the shared test_current_ci_section_matches_selected_record: the base guide names frontier6030 while the selected record requires make frontier7000-verify. The log reports 760 tests, one failure and 24 skipped; package replay was skipped. This package changes no shared CI file. Local replay remains the scientific evidence.

Nick Petalas (@petalas) supplied the signed witness and model, compact geometry and flags and unchanged checker. His source credits DreamingOfClouds (@DreamingOfClouds), with Claude Opus 5.5 assistance, for the C=56 model, chain configuration and split-gather question in PR #7689 at 511e7d0852acdca291b50c538b978fb063ae7bac. All six source files and notices are retained byte-exact. Ali Romman (@aliromman) supplied the exact optimization, proof and new checker with substantial OpenAI Codex assistance.

Changed flags, split reads, other row models, native realization, storage/restoration and the complete paid caller invoice remain open. The full PR7689 global-flag and Lean suites, selected cold replay and Docker were not run. All five framework assumptions and the selected frontier remain unchanged. No independent human/formal-kernel acceptance or global literature priority is claimed.

Base: canonical main e54523517d0f03e58352e647c403a5af3e525a57, priority revision 8. Source pins do not assert unmerged ancestry. Apache-2.0. Canonical claim.

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.

1 participant