Skip to content

Prove the sharp physical prefix range for every active tensor repetition - #7731

Open
aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-059
Open

aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-059

Conversation

@aliromman

@aliromman aliromman commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

For every positive integer m, the same ordered prefix of the fixed role-1256 graph has the largest absolute real or imaginary physical component. The sharp bound is

M(m) = (12927/2)*2^(2m) + (36549/8)*2^(4m)
       + (51071533/256)*2^ceil(9m/2).

Terminal row 156140 attains this bound after its eleventh donor 103886, in chart 127, at real output address zero. This answers the active-tensor question in PR #7725. Its one-tile result remains prior work. No new kappa or complete root-width reduction follows. The graph, paid routes and scalar factors are unchanged.

Read the proof, exact receipt, checker, source pins and credit. Each chart transition is repeated on m active factors, while scalar coefficients occur once. Every real and imaginary entrance component is independently in [-1,1] on the full 9m-axis field. Inputs need not be separable tensors.

The proof derives the exact transition gain 2^ceil(m*d/2). For rank nine and two repetitions it is 512, not the squared one-tile gain 1024. The checker groups signed prefix coefficients by transition rank, then proves polynomial dominance for both parities. All coefficients after translation by t=1+s are nonnegative. This covers all required t>=1, hence all m, without a finite repetition sweep.

Evidence Coverage
Complete ordered scan 172,414 prefixes; 9,044,092 exact coefficient products
All-parameter dominance 86,838 distinct rank-mass vectors; 1,736,760 nonnegative coefficients
Separate backward expansion All 448 nonzero coefficients of the winning prefix
Actual physical witness rows 90 entrance charts; 360 exact tensor-histogram checks through m=4
Sharp values m=1: 51863149/8; m=2: 103416050; m=3: 3287704864; m=4: 52598313856

The attaining-input formula chooses each component from the signs of the exact tensor adjoint row. The finite histogram checks corroborate that formula without materializing exponential address arrays. A reduced winning coefficient is rejected in both parity certificates. The wrong multiplicative gain and odd-rank floor are also disproved.

Python 3.12, standard library only:

python3.12 -B scripts/research_package.py --package research/relay1256-tensor-prefix-range

Both normal and optimized commands print:

PASS_RELAY1256_ALL_TENSOR_PREFIXES 172414 prefixes; all m >= 1

Registered normal/optimized replay, commit-byte/source-history/mode checks, export verification and read-only sealed serial replay pass. The sealed replay uses no Git and leaves all hashes and modes unchanged. Full predecessor suites, selected cold replay, Lean and Docker were not run. 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 supplied the one-tile theorem and certificate in the source proof. Ali Romman's PR #7709 supplied the normalized prefix method and question. Clear1nsight's PR #7704 supplied the graph and physical contract; SYLforge supplied the word. The preimage-frame source credits icekylinx and OpenAI assistance. Native and framework notices preserve all predecessor credit.

The graph hash is 8e7b755bd2d2d387bc3d3adfe48100985adf74775ca5acbb143170a3d16b5ccf. The source scientific commit is d88ecbf3dffa8a2386c90766829b769dd8068a1f, published at 72986db329dc83b77b9e69b25396c3da6d1a94bd. This PR starts from canonical main e54523517d0f03e58352e647c403a5af3e525a57. All 23 retained files are unchanged and byte-pinned. The source graph reader, backward expansion and native interpreter are shared dependencies; no pending-source ancestry is asserted.

The result covers exact row accumulators under the stated independent-input contract. Caller correlations, numerator temporaries, paid-child interiors, rounding and complete root, stock and tape costs remain open. All five framework interfaces remain conditional. This is source-aware author checking, not blind, independent human or formal-kernel acceptance. The selected frontier is unchanged.

Prepared by Ali Romman with substantial OpenAI Codex assistance in research, implementation, verification and writing. Apache-2.0. Canonical claim and handoff.

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