Repository navigation
Conversation
Replace the complex motif's per-pair side wires with a shared-sum circuit (3,693,800 -> 108,195 side roles per invocation at h=25). Binary coordinate and pair-star labels give every frame residual an orthonormal basis; injected disjoint pieces must leave an outside point uncovered, since otherwise the residual is an alternating hyperbolic plane. The certified complex saving rises from 418/10^12 to 14/10^9, so the paired bit network becomes the only binding motif. With the compact-control recurrence this gives the conditional witness kappa = 59/10^11 > 2^-31, 590/83 (about 7.11) times the preceding 83/10^12. - scripts/complex_circuit.py: circuit, F2 label algebra, role compiler, both-direction frame checks, exact dirty-scratch simulator - scripts/complex_network.py: full h=25 checks, counts, guard, gate bound, parameter witness -> certificates/complex-network.json - scripts/make_complex_circuit_patch.py -> patches/complex-circuit-31.patch - notes/complex-circuit-*.tex and artifacts/complex-circuit-note.pdf - tests/test_complex_network.py; README, status and reproducibility docs Contributed by eumemic. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The upstream workflow limits each job to ten minutes. Cache residual and label verdicts by canonical basis (same checks, entry/exit checks now also require full label validity), and let the patch generator recompute only the counts and witness instead of the full finite checks, which scripts/complex_network.py and the tests already run. The certificate drops its diagnostic count of distinct cached checks; the patch is unchanged. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Evaluate the Gaussian maps S' and E as chirped block correlations with the established multiplier: O(p^(1+delta)) per output - Bound powers of the resampling correction by a potential argument, so the Neumann series needs O(p/alpha^2 + 1/theta) terms - alpha = floor(sqrt(b/(8d))): Gaussian cost O(d^2 p^delta), epsilon may approach 1/2; beta = 19/25 fits the compressed complex saving - Conditional witness kappa = 1479/10^12 > 2^-30 (1479/590 above 59/10^11) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Feed each center wire from its group total inside the side circuit: it rises only to the star span H_i, so it loses h-1 instead of h dimensions per invocation in both stage directions - Align the paired blocks across common points and share each top-level pair-star chain between its two groups: 494196 roles vs 509244 - Bit saving 296/10^11 -> 325/10^11; witness 1479 -> 1624/10^12 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
|
Thank you @rohanarun. This contribution is now recorded on main in the scoped reconciliation review and contributor credits. Contribution: Partial-swap ternary construction removes a legacy outer shear. Review and limits: Focused tests include nonsymmetric controls; source archive retains all alternative experiments with their stated scope. The complete original source, proof/research notes, notices and relevant replay/CI evidence are retained through the pinned source archive. The review distinguishes fresh runs, exact-head historical CI and prior integration evidence. Closing this historical PR as reconciled; its obsolete headline does not replace the current selected bound, and draft/retracted claims retain their explicit qualifications. Reviewed head: |
Partial swaps on the ternary six-block network
Apply icekylinx's #18 partial-swap frames to our #17 ternary construction. This yields the conditional witness
about 11.09% above #18 in exponent saving, not practical runtime. Subsequently superseded by Zhihao Chen’s #21 (κ = 5.499e-6); this PR remains a checked research record.
The new frames remove exactly N exceptional source calls, giving
s=Wm-2N+2L. The h30 producer, physical role allocation, matching, nested basis and all six recursive block multiplicities remain unchanged. The source-relocated auxiliary exit keeps the same projector. The circuit now directly implements interchange; the legacy outer shear-conversion passes are explicitly removed.The bit saving is 4187/10^9. The unchanged h28 complex construction uses its stronger certified saving 4191487/10^12, beta=1/1000 and C1=3749/2500. All 29 constraints and seven margins are strict; the final absorption gap exceeds 6.16e-13.
Proof and validation
research/partial-swap-ternary/proof.tex.make verifypassed 308 tests and 21 inherited patch checks. Its source hashes are revalidated here.make verifyexited 0 at56d4216: all 312 published tests plus four unpublished follow-up tests passed (316 observed), and all 21 inherited patch checks passed. Physical producer regeneration completed. Seeresearch/partial-swap-ternary/validation.jsonfor scope and the log digest.The follow-up documentation correction names all 33 complex rank classes and explains the inverse early-copy split 27+729 already counted by the certificate. The full candidate witness, four tests, manuscript patch and PDF were rechecked; numerical bounds are unchanged.
Finite checks and an independent agent audit are not formal verification or independent expert mathematical review. The general basis, arbitrary-width tape, precision and inherited multiplication interfaces remain mathematical dependencies.
Attribution and review base
The partial-swap frame and triangular-conjugation identity are icekylinx's #18 contribution at
f2ab41aebad47861caf6316282c1793e5513845e; pinned unmodified sources, license, NOTICE and hashes are inreferences/pr18/. We claim this compatible ternary composition, not those underlying ideas.This branch is based on #17 (
c915b4117626a512dd470adbdbd76963108f5cb3), itself based on #15; the main-branch diff therefore includes those pending dependencies. Review the final composition commit for this addition.Credit to eumemic (#13/#15 producer, source frames and complex batching), Zhihao Chen (#16 nested basis and earlier five-subset work), icekylinx (#10 tape/batching and #18 partial swaps), Rohan Arun (#14/#17 geometry and composition), Douglas Colkitt, OpenAI and prior contributors. Prepared by Rohan Arun with substantial OpenAI Codex assistance and parallel proof/test reviews. No global-optimality or endorsement claim is made.