Community research maintained by Douglas Colkitt — conditional on the original OpenAI #109 framework.
The reviewed community witness gives
This uses the fixed finite-alphabet Turing-machine model with a fixed number of one-dimensional tapes in OpenAI's Integer multiplication below n log n. The saving is 6.10% above our previous PR #39 release, and still below 2^-14. These numbers compare asymptotic exponent savings, not practical running times.
Latest circuit contribution: Rohan Arun (@rohanarun), PR #49. This composes Chafik Boukhalfa's reordered exclusion sums and exact recovery, RaD / hipotures's alternating producers and physical compiler, and Rohan's weighted matching and order search. Their full dependency chain and AI-assistance disclosures remain credited in the source notices.
Proof and reproduction guide · Exact certificate · Maintainer review · Integration record
The community work combines recursive batching and partial-swap frames with semantic precision bounds, arbitrary-coordinate routing and bulk Gaussian resampling. Two-stage circuits, paid copied-center operations and improved contiguous blocks strengthen the finite networks. The latest increment reorders disjoint sums and retains more compatible carriers, with both local bases fixed and the entire physical circuit replayed exactly.
The selected bit network has m=575 and 177,284,805 roles. Its recursive saving is 4124034054/10^14; the unchanged complex network supplies 717/10^7. The assembly retains all seven strict exponent margins, including numerical, movement and normalization costs.
Also incorporated: Alejandro Zarzuelo Urdiales's Gaussian parity and finite tensor proofs (PR #45) and Ryan S's historical Lean certificates and circuit checks (PR #26). These strengthen verification within their stated scope; they do not change κ or formally verify the complete multiplication theorem.
This is a community result. Principal incorporated contributions include:
- Rohan Arun (@rohanarun): corner geometry, fixed-basis composition, weighted matching and the latest #49 circuit.
- Chafik Boukhalfa (@chafreaky): exact data recovery, independent checkers and reordered exclusion sums (#43, #46, #48).
- icekylinx: recursive batching, partial swaps, fixed projector profiles, copied retained centers and the selected complex construction.
- Zhihao Chen (@jacklightChen): controlled bases, translated frames, semantic/bulk compatibility and two-stage integration.
- RaD project (@hipotures): semantic precision, arbitrary-coordinate routing, phase-cell inversion, bulk resampling, alternating pair order and the independent physical role compiler (#41).
- James Chang (@jamesyc): reversed two-stage geometry and exact controls.
- Aurel Prosz (@Paureel) and Swapnil Jain: attributed two-stage development and the paid copied-stream endpoint construction.
- Dominik Scholz (@DominikScholz): dimension, parameter and fixed-basis refinements.
- eumemic: complex circuits, Gaussian resampling and source-frame work; Bortlesboat and dleen: aligned pairing, retained totals and sharing.
- Alejandro Zarzuelo Urdiales (@alejandrozu): Gaussian parity, finite tensor execution proofs and mixed-center reference checks (#45).
- Ryan S (@princezuda): historical Lean certificates, frame/movement lemmas and independent circuit checks (#26).
The full contribution record also credits parallel, incremental, superseded and pending work. Inclusion there does not claim incorporation or verification of every PR. Douglas Colkitt maintains the project and its original research, review and integration, with OpenAI Codex assistance. OpenAI's original manuscript and Harvey–van der Hoeven's analytic work retain their attribution. Contributor-specific AI disclosures remain in NOTICE.
The selected contribution is PR #49 at f95d2910e027495983b53cae1693cf535abf2569.
The follow-up review accepts its
increment conditionally on the retained PR #39 audit
and original #109 framework. This is not a claim of full formal verification,
independent human peer review, worldwide priority or optimality. No complete
practical multiplication-machine implementation is supplied.
Validation includes fresh finite producers, exact rank profiles, full dirty workspace restoration, historical patch checks and an independent rational moment/assembly checker. Both formal submissions were built with their pinned Lean versions; their axiom audits accept only Lean's standard axioms. See the review receipts and combined integration record.
The earlier PR #39 release and 2^-30 checkpoint (1a74950) remain preserved,
with all earlier certificates, proofs, patches and contributor notices.
Requires Python 3.11 or newer, Git, Make and a C++17 compiler with unsigned 128-bit integer support (tested with GCC and Clang). No third-party Python packages or network access are needed for the arithmetic/circuit verification. The separate formal targets require Lean and an initial toolchain/Mathlib download.
make verify
git diff --exit-code -- certificates patchesFor the final witness and the independent arithmetic/source audit:
make copied-fixed-verify climbed-48-verify
make community-followup-check
make formal-verifyAllow several minutes and multiple gigabytes of memory for full producer rebuilds. See reproduction details.
Each patch applies independently to the unmodified pinned source; they are alternatives, not a sequence to apply together. The result history records the earlier mechanisms and scoped ceilings.
| Patch | Conditional saving | Scope |
|---|---|---|
| frozen-154 | 2^-154 |
Original network and recurrence exponents |
| balanced-153 | 2^-153 |
Balanced assembly parameters |
| same-network-129 | 2^-129 |
Original network, sharper recurrence comparison |
| h46-111 | 2^-111 |
Smaller network, dyadic parameters |
| h46-109 | 2^-109 |
Rational recurrence saving, strict final margin |
| h46-108 | 2^-108 |
Variable stopping exponent |
| h46-rational | 5.8e-33 |
Strongest supplied parameter-only witness |
| nonadjacent-layout | Original parameters retained | Routing proof and revised layout cost only |
| frozen-nonadjacent-107 | 2^-107 |
Direct routing, original network and recurrence exponents |
| h46-nonadjacent-78 | 2^-78 |
Direct routing with the h = 46 network |
| h46-nonadjacent-76 | 2^-76 |
Direct routing with tuned dimension and stopping parameters |
| h46-shared-side-75 | 2^-75 |
Stage-1/stage-3 side-role sharing, routing, and parameter tuning |
| h46-incidence-67 | 2^-67 |
Rectangle incidence circuits, full auxiliary sharing, routing, and parameter tuning |
| h46-dag-63 | 2^-63 |
Shared intermediate sums and reversible role allocation |
| h46-shared-point | 13*2^-66 |
Cross-group sharing |
| h50-paired-59 | 2^-59 |
Paired sums, stopped guard and tighter Gaussian setup |
| compact-control-34 | 83/10^12 > 2^-34 |
Compact controls, complete reservations, local repair and separate complex arity |
| complex-compression-31 | 2^-31 |
Weighted complex circuits, binary phase frames and complete auxiliary sharing |
| ternary-30 | 2^-30 |
Ternary five-subset circuit, rational frames and fixed-alphabet interchange |
Use CITATION.cff, cite the individual contributions used and include the repository version or commit. CONTRIBUTORS.md, NOTICE and source-specific manifests preserve the dependency credits.
The project is Apache-2.0. Bundled RaD sources retain their separate
CC0 license and notices. The pinned original OpenAI manuscript remains unchanged
under upstream/; its source hashes are in upstream/manifest.json.
This project is not an official OpenAI release or endorsement.