Skip to content

Prove sharp scalar ranges for the fixed role-1256 graph - #7709

Open
aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-044
Open

aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-044

Conversation

@aliromman

@aliromman aliromman commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

This proof gives the sharp normalized scalar ranges for the fixed role-1256 graph in PR #7704. Its largest ordered row prefix is 53895757/256 for independent unit-bounded entrances. The source's sum of separate product bounds is 13819451/16. Keeping the cancellations between repeated entrance contributions gives the sharper value. No new kappa or selected-frontier change follows.

Read the proof, exact receipt and attaining inputs, checker and source credit.

Quantity Exact maximum for unit-bounded entrances Extra signed bits
Completed row or entrance 2140811/96 15
Ordered partial sum 53895757/256 18
Separate rational product 4340699/24 18
Integer numerator before division 36262265/32 21

The domain fixes all 1501 independent entrances, scalar rows, coefficients and donor order of graph 8e7b755bd2d2d387bc3d3adfe48100985adf74775ca5acbb143170a3d16b5ccf. The l1 norm of each entrance coefficient vector proves its range for every real input in the cube; the matching input signs attain it. A paid move is the identity only in these normalized scalar coordinates.

For signed W-bit integer entrances that are multiples of 921600, W+21 data bits suffice for the stated scalar arithmetic. A reachable 32-bit fixture produces numerator 2433343030560000, above the W+20 positive limit 2251799813685247. Thus the product and prefix bounds alone miss a real overflow condition. The largest prefix occurs in terminal row 156140, which eventually cancels to entrance 1372. Completed-row bounds alone also miss it.

The source's physical W+26 service has a different W contract. This result does not replace that service bound. Physical chart operators, child interiors, common-root precision, tape implementation and compiler cost remain separate. All five selected assumptions remain; the selected kappa stays 0.00363005817. The full source endpoint theorem, caller invoice and physical application are not certified here.

The base is current main e54523517d0f03e58352e647c403a5af3e525a57. The consumed graph is from @Clear1nsight's exact head e29a5ae8591243c1178067727f12ce4abefa0b93: graph, physical analysis, source checker, source receipt. The graph bytes, license and original notices are retained. The source verifier is not imported.

Credit @SYLforge for the nine-swap word in PR #7700, as recorded in the source notice. It retains the contributions of DreamingOfClouds, Ali Romman, Rohan Arun, Jerry Chen, Douglas Colkitt and the bank/caller/framework lineage. Ali Romman's PR #7702 supplies related row-atomic arithmetic on a different graph and donor contract; its bound is not consumed. Prepared by Ali Romman with substantial OpenAI Codex assistance, Apache-2.0.

With CPython 3.12 and the standard library, run:

python3 -B scripts/research_package.py --package research/relay1256-sharp-scalar-range

Both normal and optimized commands print:

PASS_RELAY1256_SHARP_SCALAR_RANGE
7644 rows; 172414 prefixes; 9044092 basis products; 8 scalar oracle runs
Extra signed bits: snapshots 15; prefixes 18; products 18; numerators 21

All 172414 prefixes and 9044092 nonzero basis products are checked. Eight complete scalar oracles attain the four maxima over rational and integer inputs, with four one-bit-smaller failure witnesses. Source bytes, chronology, charts, exact divisions and the source triangle bound are checked. Local normal/optimized replay, committed-byte/mode and source-history checks, export and sealed Git-free 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. Complete predecessor suites, selected cold replay, native/physical replay, Lean, Docker, human and formal-kernel review were not run.

The scoped claim takes the scalar-range sub-obligation. Live source/phrase overlap and the closed delta were checked. The concurrent sender/receiver work is now published as PR #7708, current head 7017edbe4bc5ed1b6d5b6defedbc7e472d83c65c. It changes the graph and keeps a no-cancellation common-root guard. This audit retains the graph supplied in PR #7704; its exact numerical bounds do not transfer automatically to #7708. It changes no row or invoice. Sparse local and unpublished work remain overlap limits. The README states the open physical-promotion questions.

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