Repository navigation
Conversation
2992f23 to
6d92445
Compare
ahrazzle
left a comment
ahrazzle
left a comment
There was a problem hiding this comment.
Target: PR #433, head 6d92445
(base 12a4934, fetched 2026-10-11T12:13:32Z).
Verification
Independent third-hardware replay of this head, from a fresh detached
checkout at the staged SHA.
- Verifier exit 0 in both modes. Default mode reuses hash-matched phase
evidence and passes 13/13 steps. Full mode (--replay-phases) regenerates
every selected Gaussian phase pair and passes 13/13 steps with 2918/2918
exact phase edges. - A second independent replay on separate hardware also returns verifier
exit 0, 13/13 steps PASS, 12.33 s wall. - kappa = 560432476129139/500000000000000000 = 0.001120864952258278.
The fraction is exact to the last of 18 decimals and identical to
EXPECTED.json, CERTIFICATE.json and this PR's body. - Binding component: bit. The bit lower bound sits below the complex lower
bound by exact cross-multiplication, so the bit moment bound is what caps
kappa. - Recomputed gain over the #425 baseline is 6.9522005058117555%, matching
this PR's 6.9522%. - Digit-exact: an independent closed-form re-derivation (stdlib fractions,
no package code) reproduces the same fraction on the 1e-18 grid. The same
code reproduces the earlier published 203592481059369/200000000000000000
exactly, so the formula is not fitted to this PR.
Negative controls
All rejected. The package assembly controls next_grid, zero_eta,
zero_beta and unit_beta each raise. So do 8 bit mutations, 7 complex
mutations and all 18 bridge-shear omissions. In an independent perturbed
family, dropping the binding bound by one 1e-18 tick makes the claimed
kappa fail at the first constraint. The accepted margin is 0.585 of one
tick, so the certificate is tight rather than slack-passing.
Limits
The claim is conditional as this PR states: the all-size compiler,
compressed-window, correlated dirty-bank, restored-work, tape and common-grid,
precision and analytic interfaces remain assumptions, and no Lean-kernel
proof or unconditional multiplication theorem is claimed or verified here.
AI-assistance disclosure: prepared with AI assistance under human oversight.
…, wording)
- check_extended_labels.py: a source must convert exactly once (entry adapter
only from ('pre', i)); identity adapters pinned to frame 0 / FULL; wrappers
keyed by the start frame; unknown events rejected; new mutant (duplicated
source entry adapter); adapter mutants now also run on the anchor.
- code/check_adapters.py (new): #433's data-adapter identities on all 512
columns of every port (H6), 0 failures on all four words.
- verify.py --instances: confirms #484's own checks passed on inputs equal to
#484 EXPECTED.json, runs both instance checks; #484 files it imports are
pinned in MANIFEST (external_files).
- PROOF/README: explicit H2 bridge and C-linearity, H6, exact statement of
#484's endpoint imprecision (mid-stream adapters are identities), "every
input" qualified by its inherited facts, semiring wording.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Model: units are physical streams (family, cover class) as in #442/#458 and the template's route_semantics. New LemmaA/Streams.lean: streamOwner_injective(_of_disjoint), sweep_unique, lemmaA_streams (one owner per stream from disjoint footprints on shared families); plus lemmaA_piece and lemmaA_async (Lemma A feeding #505's window variant). - code/check_owners.py rewritten for stream units (data wires alone, disjoint footprints on shared families, centre copies in distinct private slots, atomic calls; 3 mutants): all four #484 words pass, including the 32 gauge80 / 16 kernel79 slot-sharing calls and the anchor. - code/check_windows.py (new, independent): every hosted block inside its pending child's open window, blocks disjoint, family-group windows disjoint (2 mutants); all four words and #505's emitted async template. - verify.py: runs Lemma G's check_extended_labels.py ((F) is read in the extended labels), fails on a missing LEMMA-A.json, checks I-helper counts against EXPECTED, prints the owner and window results; --async T.json.gz. - Docs: unit model stated, coverage per word, #505 replay results and what is not re-run, commutative-semiring wording, credits #433/#442/#500. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thank you @icekylinx. This contribution is now retained on main in the scoped 100-PR review and contributor credits. Contribution: Split complex centres and bit source-basis reuse. Review and limits: A structural improvement in both suppliers, with dirty response, reflected chronology and larger prefix charge. Retain mutable-source/compiler and analytic hypotheses. The complete original source, notices and relevant execution evidence are retained through the pinned source archive. The review distinguishes fresh local runs, source-bound historical Linux CI and prior exact-head review; it does not claim 100 fresh full mathematical audits. Closing this historical PR as reconciled. This preserves parallel and cumulative credit without replacing the selected frontier or promoting priced research targets to completed constructions. Reviewed head: |
- The producer's lower_child check (signed Z4 tableau plus column 0) is complete but not independent; K2 is a control on this package's own comparator, not a gap in the producer's check. Lemma G's I-tr and check_adapters.py ran on four other words. - Scope is record level for CLIFFORD-RECIPES.json; identity adapters are 866 per orientation (352 entry + 514 exit). - Gate-order equivalence holds on all 224 adapters (counted in RESULT). - B3 accepts only the selected record or checkpoints/frontier1094.json; B8 uses a path pin in the source manifest. - Tests (26): mutated child/adapter records rejected by evaluate(), receipts cover each pair once, measured digit bound on the 16-gate identities and a beta = 2K collision, I5/I6/I7 and fixture-pin negatives, calibration reproduced. - Credits: HyperplaneEnd Lean (jeremykothe2, CrocSwap#593), Lemma G (DreamingOfClouds), CrocSwap#433 H6 identities, PR425 (jacklightChen), CrocSwap#1029/CrocSwap#662 engines.
3 participants
This combines a new split-centre complex construction with source-centre reuse on the bit side, in one commit based on main at
12a4934d15006e8c10f9b2264f13f8f341be802b.Under the retained compiler, compressed-window, arbitrary dirty-bank, restored-work, fixed-tape/common-grid, precision and analytic interfaces,
The combined saving is approximately 6.9522% above #425. The bit supplier binds the new assembly.
Construction
On the complex side, cancel the coordinate-source contributions that appear with opposite signs in the centre and side decoders before lifting them to address operators. The 112 code-line targets then use eight rank-eight centres; the eight coordinate targets use rank-one identity copies. Their total copied rank remains 72. Removing 112 coordinate-source helpers and eight dedicated full-centre helpers reduces R from 680 to 560. The dirty response is recomputed from the new word, and both orientations include literal source/helper cleanup and exact heterogeneous-centre phases.
On the bit side, start from #426's R767 word. Form its nine retained sums with an invertible basis change on the source rows, copy them sequentially, and apply the literal inverse before continuing. This removes nine helpers. Compose the resulting R758 word with the compressed dimension-64 geometry from #386, replaying the actual reflected copy chronology and all new bank owners. The changed prefix needs 223 bits and is charged at 256 bits; the old 192-bit allowance is not reused.
Both exact moments include the retained bad-branch allowance. Six fixed bootstrap levels and all 47 strict outer inequalities give the displayed κ; the next grid point and invalid-parameter controls are rejected. The result is calculated from the two emitted programs, not by adding independently reported gains.
Contribution sources and prior PRs
e11ccea1d092beb1b01b4ee4ce49da478234254c88d552526893cd15325948bb1487e9a4d4bf32f6My earlier contributions remain historical context for this work:
The new split-centre construction, source-basis composition and integration are contributed by icekylinx, with substantial OpenAI Codex assistance. Earlier Pro exploration used GPT-6 Astra. Existing authorship, licenses and source-specific AI disclosures are preserved in
NOTICE.md,notices/andSOURCES.json.Proof and validation
The selected package is
research/split-centres-global/:PROOF.md: split-centre cancellation, arbitrary-dirt lifting, source-centre/reflected chronology and conditional assembly.CERTIFICATE.json: both component bounds, fresh finite invoice, bootstrap cutoffs and all 47 strict outer margins.EXPECTED.json,SOURCES.json,MANIFEST.json: exact result, source pins and byte-level bindings.verify.py: standard-library reproduction; dedicated CI covers Python 3.11 and 3.14.The clean-directory incremental replay passed locally on Python 3.14.8:
To regenerate every selected Gaussian phase pair, add
--replay-phasesand use another fresh output directory. The default reuses phase evidence only when the regenerated canonical word and full recipe hashes match. The unchanged historical full suite was not rerun; no PDF was built. CI for this commit is not represented as passed before it runs.This remains a finite conditional construction. The inherited all-size interfaces remain assumptions; no unchanged legacy gx/gxcore acceptance, new Lean proof, global optimality or unconditional multiplication theorem is claimed. The main selected certificate is left for maintainer review.