Repository navigation
Post-center rank-four kernels for conditional κ = 0.001584931111785787 - #535
Merged
Mister-Meeseeks merged 1 commit intoOct 12, 2026
Merged
Conversation
icekylinx
force-pushed
the
research/post-center-rank4-global
branch
from
October 12, 2026 02:20
0d343e2 to
25f9b7c
Compare
icekylinx
force-pushed
the
research/post-center-rank4-global
branch
from
October 12, 2026 02:46
25f9b7c to
8b758a1
Compare
This was referenced Oct 12, 2026
This was referenced Oct 12, 2026
This was referenced Oct 12, 2026
Merged
Contributor
|
Integrated on Your post-center gauge/kernel ordering, rank-four construction and physical packing are selected on main with DreamingOfClouds’ freshly replayed #510 supplier: κ = 0.001584931111785787. The review additionally lowered all 2,927 local edges and 224 endpoint adapters, checked all 21 wrappers, local and owner mutations, and bound the fresh bit invoice/profile to the 47-constraint combined assembly. Maintainer review and exact scope · Contributor record. The multiplication claim remains conditional on the inherited all-size/realization/analytic interfaces; these checks do not establish a complete unconditional theorem. Thank you for contributing this work. |
Mister-Meeseeks
pushed a commit
that referenced
this pull request
Oct 12, 2026
…01587467473762886 A bit word on #493's D4 quad skeleton with three registers per cell. Each target reads its single d-line from a source register at a cell; after the reads, that register climbs to its block's D4 frame and feeds the block node that makes the Q-deliveries for the complementary block. R = 420 (best previous bit word: 480). Priced by #510's nine-stage chain: c = 0.001653826988146798944268; by #493's eleven-stage chain: c = 0.00164031632456626881922. Complex-bound: kappa = 0.001587467473762886 with #535's complex supplier, and unchanged 0.001517760977820331 with merged #531's. Review point: 597 phase-B gates write source registers after targets have read them; the package states #425's scalar-interface lemma with interleaved z-gates and checks each hypothesis. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Mister-Meeseeks
pushed a commit
that referenced
this pull request
Oct 12, 2026
…bit c = 0.00165947204563779652541 Adds research/e8-punctured111-star: an E8 bit word with 111 live ports (9 of the 120 labelled lines punctured), nine rank-7 pseudo-target totals (decoder constant 63), R = 489 helpers, and the parameterisation of PR #493's hosted-audit chain that prices it (55 exact string edits, one configuration derived from the word). Bit c is 4.54% above the live bit c of PR #510; with PR #535's complex b, kappa = 0.001587467473762886 (complex-bound, +0.16% over #535). Regression: PR #503's R480 word is byte-identical through the stock and the adapted chain. Eight mutation controls, all rejected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 12, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This extends the merged #531 supplier with post-center compensation and rank-four response kernels. It is one new commit based on main at
f19cb621be2caf3a7291a37c343c1c55afe935e0.Under the retained all-size compiler, common-label/mutable-basis realization, normalized Lemma A, whole-call hosting, restored-work, precision/tape and analytic-transfer interfaces,
The saving is 4.42561% above #531 and 3.98392% above #533. These percentages compare exponent savings, not practical runtime.
Construction
The 112 ordinary-M helpers remain untouched through the source prefix and center cut. Their initial dirt can therefore be compensated at the beginning of delivery. In each of fourteen disjoint groups, four response columns span the other four, and every helper's first native frame contains a common rank-four frame.
Perform the gauge compensations first, then inject each pivot into its donors at that rank-four frame. A donor can now serve both its own gauge and the larger kernel: its compensation has already read its initial dirt before the new pivot contribution arrives. Exact response relations, target chains, nested frames and the full inverse cleanup are checked on the emitted word.
The entrance inventory is 40 rank-one kernel pivots, 112 rank-one gauges and 56 rank-four pivots. The 504 residual widths are consequently 56 of width five, 152 of width eight and 296 of width nine. A fresh whole-owner assignment and finite invoice produce the global profile.
The complex supplier now exceeds the previous bit endpoint. The final composition therefore adopts the published #510 R480 nine-stage bit supplier at
0.001587447106092548137108; the bit side binds. This bit improvement is adopted and credited, not claimed as new work here. The unchanged #493 bit supplier gives the separate control κ0.001562995248531632.Both component moments are repriced by two rational engines. Six bootstrap levels and all 47 strict outer inequalities certify the displayed global κ; adjacent-grid and invalid-parameter controls are rejected.
Contribution sources and prior PRs
e07a8a3f9f79c22113398a0c5dbbc7601f489b3e64edee18af02aab0c1696b677a0287dd0146381bMy #523 supplied the 112-port restriction; #433 supplied split centers and source-basis reuse. Earlier contributions in #10, #18, #24, #32, #36, #104, #115, #130 and #144 concern batched boundaries, partial swaps, gauges, certified assembly and frame/cover constructions. These are provenance, not additive numerical gains. The earlier packages retain Dugongue's asynchronous scheduling and other source-specific credits.
The new ordered gauge/kernel composition, rank-four construction, physical integration and carryover exposition are contributed by icekylinx, with GPT-6 Astra assistance. Local review and packaging used OpenAI Codex.
NOTICE.md,SOURCES.jsonand the separate original #531/#510 notices preserve the contribution boundaries.Proof and validation
The incremental package is
research/post-center-rank4-global/; it reuses the baseline already present in main.proof/POST-GAUGE-RANK4-PROOF.md: ordered compensation, exact relations, frames and cleanup.proof/GLOBAL-HOSTING-PROOF.md: actual owners, paid profile and finite accounting.proof/PR510-MUTABLE-SOURCE-CARRYOVER.md: role-independent common-label argument and explicit bit dependency scope.evidence/audit/EXACT-ASSEMBLY.json: both component moments and all strict assembly margins.BASELINE-PINS.json,MANIFEST.json,LOCAL-VALIDATION.json: inherited inputs, submitted-file bindings and incremental validation.Local validation also used
--rebuild-local --delta-phaseswith Python 3.14.8 and NumPy 2.3.5. The full changed-word construction, scalar/frame replay, physical owner audit, fresh invoice and exact assembly passed. The newly introduced intermediate phase cache was checked against #531 (194 changed edges), followed by the final 126-edge delta. Independent local replay covered both orientations on all 728 columns; all 21 residual wrappers passed, including global phases, with 11 local and five wrapper mutation controls rejected.The manifest was generated from staged Git files, excluding all logs. A clean archive of those staged files passed the same incremental reconstruction in 14.85 seconds. Every baseline pin also resolved and matched inside that archive. The PR510 data copy differs from its pinned source only by an appended terminal LF; both byte hashes and the unchanged JSON data are documented in
SOURCES.json.Package manifest SHA256:
23143f650947a2b95b443a7ea5ded79d1d5644783b7f8a54967ba137b8b801f8.Global template content SHA256:
b9f4b60e56e5153c1fc89370e9ceb09f24b23e1c8ff56f2b47ef9d867aa04f2c.Dedicated Python 3.11/3.14 CI covers reconstruction, changed phases and local wrappers. Remote results are pending. No predecessor cold replay, full bit-operator rebuild or PDF generation was performed.
This remains a finite conditional witness. The cited #510 physical bit operator and the inherited all-size/analytic contracts remain dependencies; the package does not claim an unconditional multiplication theorem, new Lean formalization or unrestricted optimality.