Repository navigation
A stronger complex supplier: PR200's reuse pairing on the source-assisted v4 word, complex saving 7.0991·10⁻⁴ (+1.28% over #194; no new κ, bit binds) - #233
Conversation
…isted v4 word, complex saving 7.0861e-4 (no new kappa) Adds research/source-assisted-v4-layer and one workflow. PR194's pipeline (PR184's source-assisted word on the PR168 v4 modules: aligned rewrite, frame flow, exact lift, contract, assembly) is kept unchanged except the physical layer the flow receives: PR200's descent seeded from PR168's frames changes 5,577 of 17,789 operation frames and recomputes the 2,310 reuse pairs. The layer is admitted by PR200's complete complex checker on the aligned word; the flow, exact lift and contract certify complex saving 3543063/5000000000 = 7.0861e-4 (PR194: 7.00918e-4, +1.10%). With PR200's bit supplier the assembly gives kappa 6768823/10^10, unchanged (bit binds). The 308 pipeline files of PR202 (8d8d67b) are vendored with sha256 pins. Prepared by Chafik Boukhalfa with substantial Anthropic Claude and OpenAI Codex assistance.
… DFT PR CrocSwap#233 keeps PR CrocSwap#194's source-assisted word and replaces only its physical layer by PR CrocSwap#200's descent; its certified complex moment root 7.0861e-4 transfers as before (Proposition D). Vendored certificate, pin, moment certificate and controls added; expected.json regenerated.
… certificate was keyed by temp paths)
…d-artifact hashes go to report.json
…ofiles carry floats); report field-level differences
… = 7.0977e-4 (+1.26% over PR194), 435 frames changed One light descent round from PR168's frames with a 98% movability mask (seed 8) prices higher under the flow than the full cycle; certified through the same chain (PR200 checker, flow, exact lift, contract, PR194 assembly).
chafreaky
changed the title
|
Update ( |
…iring; complex saving 3549537/5000000000 = 7.0991e-4 (+1.28% over PR194) The gain is the reuse pairing, not the frames: with PR168's operation frames byte for byte and pairs_mw.py's maximum-weight late compensated pairing the flow root is 7.09907e-4; single frame moves never help (flow_moves.py screen). Same verification chain; docs rewritten for the pairs-only claim.
chafreaky
changed the title
|
Update ( |
…ring only, frames unchanged)
…7211262868e-4 The package vendors every input it reads by sha256 -- the rung-1 package of CrocSwap#219, the frontier certificate of CrocSwap#207 and the unmerged source-assisted v4 row of CrocSwap#233, its earlier heads included -- and re-derives the whole complex-side ladder offline: both paid moments per rung, the 47-constraint assembly with adjacent-grid rejection, the padded width-66 schedule, the three export contracts and their import gate. On CrocSwap#233's row the first whole-bank family of the ledger is rank 3, not rank 16: 66 = 22 * 3 makes its bank exactly filled, so the T1 padding obligation does not arise, and the point prices 1819302815717/25*10^14 = 7.277211262868e-4, bit-bound, above the ladder top and above the pinned frontier. The rung is row-dependent and that is measured: its stability is re-taken on all six ranked heads of CrocSwap#233 and all nine width-66 rows of this repository, and the rung above it is a recorded negative. Verification is wired into the repository: a Makefile target in the aggregate, its own CI group in the Python matrix, an exact-bytes .gitattributes carve-out so committed digests survive a clean checkout, and a test module whose pin controls check the digest gate fails closed. Generated with Codebuff 🤖 Co-Authored-By: Codebuff <noreply@codebuff.com>
…d prices at theta = 1 - 7547/10^7 (priced transfer), PR CrocSwap#194's at 7474/10^7 (Lean figure recovered)
…7211262868e-4 The package vendors every input it reads by sha256 -- the rung-1 package of CrocSwap#219, the frontier certificate of CrocSwap#207 and the unmerged source-assisted v4 row of CrocSwap#233, its earlier heads included -- and re-derives the whole complex-side ladder offline: both paid moments per rung, the 47-constraint assembly with adjacent-grid rejection, the padded width-66 schedule, the three export contracts and their import gate. On CrocSwap#233's row the first whole-bank family of the ledger is rank 3, not rank 16: 66 = 22 * 3 makes its bank exactly filled, so the T1 padding obligation does not arise, and the point prices 1819302815717/25*10^14 = 7.277211262868e-4, bit-bound, above the ladder top and above the pinned frontier. The rung is row-dependent and that is measured: its stability is re-taken on all six ranked heads of CrocSwap#233 and all nine width-66 rows of this repository, and the rung above it is a recorded negative. The change adds three things and edits nothing: the package, a test module whose pin controls check the digest gate fails closed, and the package's own workflow. It deliberately does not take a Makefile target in the aggregate, because `Makefile` and `.gitattributes` are recorded by other packages' manifests (the joint-dual and round-8 certificates among them) and editing either moves their digests; the workflow runs the same commands on 3.11, 3.13 and 3.14 and then requires `git diff --exit-code`. Generated with Codebuff 🤖 Co-Authored-By: Codebuff <noreply@codebuff.com>
|
Two follow-ups on this word: #256 exports it as one explicit numbered program in Jacob Sussman's gcert/1 format (the "literal globally renumbered operation program" #194's contract leaves unexported), accepted by his reference checker with the exact scalar identity, with #194's word as the control (it reproduces his published histogram of #193 exactly); and jacobalansussman/wht-power-saving-lean#2 puts it through his Lean kernel checks: Walsh–Hadamard at 1 − 7547361/10¹⁰ and the Fourier transform of every length at 1 − 7547360/10¹⁰ ( |
|
Thank you @chafreaky. This contribution is now retained on main in the scoped 100-PR review and contributor credits. Contribution: Reuse pairing for the source-assisted complex word. Review and limits: Prior exact-head full original complex verifier passed; separate bit and outer dependencies retain their explicit scope. 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: |
2 participants
Adds
research/source-assisted-v4-layer/and one workflow; no existing file changes, no new κ. PR #194's pipeline (icekylinx's #184 source-assisted word on the PR #168 v4 modules, ikeboy's rebuild: aligned rewrite, frame flow, exact lift, contract, assembly) is kept unchanged except the 2,310 reuse pairs of its physical layer; PR #168's operation frames are kept byte for byte. With PR #200's maximum-weight late compensated pairing (pairs_mw.py) in place of #168's pairing, the certified complex saving becomesComposed with #200's bit supplier through #194's unchanged 47-constraint assembly, κ = 6768823/10¹⁰, the same as #202: the bit branch binds (effective 6.7734·10⁻⁴). The complex cap of every composition on this lineage rises from 7.0092·10⁻⁴ to 7.0991·10⁻⁴, which is the budget the banked bit rows (#205–#223) and the conditional rungs (#219/#224) are priced against; the exact-DFT exponent saving of #225 rises with it (follow-up there).
What changes
The flow's child histogram depends on the reuse pairs it receives: a pair identifies a dead ungauged donor's register with a gauged recipient, and the equal-frame compression follows those identifications. #168's pairing is a maximum matching of donors to recipients in a fixed order; #200's is also a maximum matching with the same legality (donor ungauged and non-root, dead before the recipient's first operation, end frame inside the recipient gauge) but a different selection among the maximum matchings (transversal-matroid greedy with augmenting paths, donor weight the gain heuristic). Nothing else changes: frames, matching arcs, modules, local circuit, gauges, aligned rewrite, flow options; the kernel pairs the flow finds on this layer are recorded as #194 records its own.
physical_opt.py:pairs)pairs_mw.py)Random pairings fall to 6.4–6.6·10⁻⁴. Descending the frames as well does not help: a screen over every single endpoint move on top of #168's frames finds no improving move, and the full descent cycle gives 7.0861·10⁻⁴ (an earlier commit of this PR, since superseded). The profile keeps #194's W = 12,052, rank mass 794,112 and deficit 1,320; only the child histogram changes.
What is checked
python3 -B research/source-assisted-v4-layer/verify.py(about 80 s; numpy/scipy for the regeneration steps; -O refused):baseline-pr202.tar.gzcarries the 308 files of Compose the v4 source-assisted complex word with the PR200 bit supplier: conditional κ = 6.768823e-4 #202 (8d8d67b) the pipeline runs (Source-assisted v4 complex word with PR189's bit supplier: conditional κ = 6.674324e-4 #194's package, Source-assisted common-frame compression for conditional kappa = 6.566436e-4 #184's decision/global scripts, the Annealed modules, terminal sinks and physical bit words on paired cubes: conditional κ = 6.558894e-4 #168 v4 scripts, references and inherited witnesses, the Clifford frames note, A stronger bit supplier: fixed-coordinate and face-diagonal bit local circuit on PR168 v4 (effective bit saving 6.7734e-4) #200's certificate, arithmetic and complex checker), each pinned by sha256 inSOURCE.json(git show 8d8d67b:<path> | sha256sum); the three witness files are pinned too.complex/physical.py+verify_fused.py) re-executes the aligned word with the package's pairs: node spans inside nondegenerate frames, nested role chains, every alias read at its deadline after the donor's last operation with the donor's end frame inside the recipient gauge, target read chronology, exact adjoint, and every formal source/target/dirty column in both shear signs on an injective packed basis with an independent coefficient bound.assemble.pywith A stronger bit supplier: fixed-coordinate and face-diagonal bit local circuit on PR168 v4 (effective bit saving 6.7734e-4) #200's bit certificate reproducescertificate.jsonexactly (exact fields only; the full run record isreport.json); the next grid point is rejected by the unchanged assembly.Run before opening: the verifier passes locally (Python 3.14, macOS, numpy 2.3.5, scipy 1.17.0) and in this PR's workflow on Python 3.11 and 3.13.
Scope and limits
Conditional exactly as #194: the retained analytic, uniform-recursion, precision, weighted bit frames, fixed-tape routing, restored-row and all-size semantic interfaces remain hypotheses; the scalar replays and exact lifts are finite checks. #184's independent complex scalar transcript is not re-exported; the one-column identity is replayed by #200's checker on the pre-flow word and by the lift's inverse gate programs. The pairing is a discovery heuristic (
discovery/, not run by the verifier); no optimality among maximum matchings is claimed. Nothing here changes κ.Credits
icekylinx (#144 paired cubes, compiler, frozen matching, gauges, Clifford frames; #184 source-assisted construction, frame flow, exact lifts, contract, assembly; GPT-6 Astra and Codex assistance), ikeboy (#191/#193/#194 rebuild on the v4 modules and the composition), eumemic (#117/#155/#157/#161/#168: modules, physical frames and reuse, output merging, annealed modules, the physical layer whose frames are kept here; Claude and Codex assistance), James Chang (#124/#166), DaysSky (#162/#192), GamingPuzzled (#160), gupt1156 and Abhinav Ramachandran (#148/#158), Zhihao Chen, Rohan Arun, an664, and the earlier lineage in the repository NOTICE. #200's pairing, checker and bit supplier and #202's pinned tree: Chafik Boukhalfa. Full attribution in the package NOTICE. Prepared by Chafik Boukhalfa with substantial Anthropic Claude and OpenAI Codex assistance; Apache-2.0. This package claims only the pairing, its admission on the aligned word and the resulting certified complex saving.