Skip to content

Prove the all-column native operators of the complete grade-14 multiplier - #7837

Open
aliromman wants to merge 2 commits into
CrocSwap:mainfrom
aliromman:codex/complete-claim-assumption
Open

aliromman wants to merge 2 commits into
CrocSwap:mainfrom
aliromman:codex/complete-claim-assumption

Conversation

@aliromman

@aliromman aliromman commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

This proof removes native_all_columns from the complete grade-14 multiplication claim. For arbitrary binary operands of at most n bits, the fixed-tape time remains O(n (log n)^(1−κ)) at κ = 0.00387992802601. The original application needed five premises. This construction needs four. The limiting outer cost remains the main transform. Operator proof · Physical routes and leaves · Complete application · Exact receipts

The eliminated premise required the actual operators, all phases, spectators and whole-job adapters to work at every positive column multiplicity. The native one-column replays and grade counts did not prove that implication. This proof starts at the two words consumed by PR #7835, follows arbitrary independent inputs through both recursive boundaries and supplies its imported ordinary interchange rate from elementary operations. Original five statements · Exact word, phase and price inputs

Interface Construction and verification
Complex Regenerate 2,955 relative Clifford jobs and 208 absolute-phase boundary monomials. Check all 1,289² rational transfer coefficients in each direction. Lift the native chart changes across columns, while retaining each caller coefficient once. This realizes all 48 calls and 30,000 global normalizers at their existing prices.
Binary Use each actual port's head line in the 134-axis caller. Lower the complete low/high cover permutation, its rational carry, the matched donor and inverse return, and both good and nonunit weighted leaves. All 36 calls yield the original 20,706 productive outputs and 107,460 paid positions.
Ordinary input and integer output The elementary digit exchange funds one preparatory grade-14 stage. Its envelope 0.003864408313999584 exceeds the required seed 0.00385248. The four original priced stages then deliver 0.003879928026095. Both supplier rates and all nine outer cost rows support the original κ.

Remaining dependencies. The four original premises are owned_erasure for private streams and suspended owners, compiler for finite chronological compilation and uniform closure, precision_tape for the fixed-tape movement, storage and precision prescriptions, and analytic_recovery for the outer decomposition, prime avoidance and exact recovery. They keep their original statements. The proof is specific to the grade-14 application and the words and weights it actually uses. The conservative constants are enormous. We did not materialize either full uniform template or execute an asymptotic-size integer product. Independent human and formal verification of the universal proof remain open. Qualification and comparison with higher advertised rates

Verification. CPython 3.12, standard library only. Run from the repository root.

python3.12 -B scripts/research_package.py --package research/grade14-native-realization

The five scientific checks emit these markers.

PASS_GRADE14_ALL_COLUMN_COMPLEX
PASS_GRADE14_ALL_COLUMN_BIT
PASS_GRADE14_COMPLETE_LOCAL_RING
PASS_GRADE14_FOUR_PREMISE_APPLICATION
PASS_NATIVE_ATLAS_BUILD c7173d46c95d1311b8b18a92c874fb7013c6e694aa8e9c0c256f0f1eaf1ede74

We reproduced the three original grade-14 receipts byte for byte. The new registered replay and the read-only, Git-free export replay passed. The independent physical checks cover 1,930,240 monomial basis images, 95,420,416 Walsh phase values and all 108 split rank-one projectors over Z/9 at four widths. Replacing all binary port placements by identity changes 27,972 basis directions. Dropping the mixed-role inverse return changes all 7,776 positions in that control. These finite checks support the written arbitrary-width proof. Complex verification · Binary verification · Independent leaf check · Application verification · Replay manifest

GitHub's hosted onboarding test stopped the plan before package replay. The canonical CI guide still lists the historical 6030 replay where the selected record requires make frontier7000-verify. The same test fails on the unmodified canonical base. All six files it reads are byte-identical between that base and this PR. Direct package selection on the exact change selects only this package. Both local five-command replays passed. Hosted failure · Canonical navigation test · Canonical CI guide

Base and credit. The canonical base is e54523517d0f03e58352e647c403a5af3e525a57. @rohanarun submitted the grade-14 cutoff at 43007c979859fa84ccc3a3509ccacc9620d181fa, following @Dugongue's preceding cutoff. @Clear1nsight supplied the actual rational word at 14c74c6a67108f675ac1196a5f2415ead0920e71 and the binary word and expanded verifier at 89a67c404226d51028e135de2b9169e9020d57af. @aliromman supplied the previous mixed caller at 4f7c67ae387cc91580f469aeda4d508b372b9ec1. @CubeJerry and the credited predecessors supplied the selected circuit and Clifford engines. PR #7831 · PR #7830 · PR #7786 · PR #7776 · PR #7000

Ross Richardson supplied the fallback implementation, Jose Manuel the recursive ordinary closure, and Nick Petalas the complete-cover counting argument. The operator proof uses the earlier monomial lowering, fresh-row and delayed allocation proofs, and the retained compact global router. Ali Romman prepared this proof, implementation and source-bound checks with substantial OpenAI Codex assistance, including separate adversarial reviews of the complex, binary and resource arguments. The source notices credit the rest of the physical, compiler and analytic lineage. Full credits · Canonical proof and code pins · Coordination and review request

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