Skip to content

Repository files navigation

Certify: replayable certificates for machine-checked mathematics

Archival record (v0.1.0): doi:10.5281/zenodo.21799112

Author: Daniel Kirtchakov (Independent researcher, Half Ounce Research) — daniel@halfounce.io. Repository: https://github.com/05oz/certify. Date of this snapshot: 2026-08-04.

Two independent bodies of work live here, sharing a method rather than a subject: produce an artifact a skeptic can re-check without trusting the tool that made it, then say exactly what still has to be believed.

Part A — Alpöge Keller map (v0.1.0) Part B — quantum code distances (v0.2.0)
Subject Degree minimality in the equivariant class of the Alpöge Keller map; moment-map structure of its cotangent lift Certified minimum distances of eleven stabilizer codes, through IBM's [[288,12,18]]
Artifact 8 msolve Gröbner unit-ideal certificates + SymPy verification scripts witness pairs + LRAT unsatisfiability proofs + ZX-duality permutations
Checker scripts/ (SymPy) qec-scripts/ — 649 lines of Python, standard library only
Paper paper/preprint-dixmier-poisson.* paper/preprint-qec-distances.*
Certificates certificates/ qec-certificates/
Provenance PROVENANCE.md §§1–4 PROVENANCE.md §5, SWEEP-RECORD-QEC-2026-08-04.md

Every mathematical claim in either preprint maps to a script whose assert statements pass, or to a stored certificate a checker accepts. Nothing is conjectural unless labeled so.


Part A — the Alpöge Keller map

Degree minimality in the equivariant class of the Alpöge Keller map, and the moment-map structure of its cotangent lift — certificates, verification scripts, and preprint. All computations carried out with Claude (Fable 5), SymPy 1.14, msolve 0.10.1. Everything here is in characteristic zero.

Credit — read this first

  • The map. The degree-7 Keller counterexample F to the Jacobian conjecture in dimension 3 is Levent Alpöge's, announced 2026-07-19. The problem was posed by Akhil Mathew; per the announcement, the search that produced the map was run with Claude Fable 5. Everything in this repository is downstream of that example. See T. Tao's expository post (terrytao.wordpress.com, 2026-07-21) and the Secret Blogging Seminar thread (sbseminar.wordpress.com, 2026-07-20).
  • The Dixmier corollary AND the symplectic cotangent lift. Both are W. G. P. Mayner's, from the same repository and the same day. The statement that DC_n is false for n ≥ 3, with the explicit Weyl-algebra endomorphism Ψ_F as witness, was first stated publicly in a Secret Blogging Seminar comment (2026-07-20, 22:08), backed by a GitHub note prepared with Claude Fable 5 (github.com/wmayner/dixmier-counterexample, two commits on 2026-07-21). The accompanying REPORT.md §8 in that same repository contains the cotangent lift Φ(q,p) = (F(q), G(q)p), machine-verified there to preserve the standard symplectic form exactly, to have det DΦ = 1, and to be non-injective; his priority section §12 lists "the symplectic lift to C⁶" explicitly. We credit both fully and claim neither. An earlier version of this README credited Mayner for the Weyl endomorphism only — that was an attribution error, now corrected.
  • Most of the structural material here was anticipated. The master equation, the anchor lemma, the parked square, the S₃/discriminant computation, the trace identity, and the exact image theorem were all obtained independently here and then found to have earlier public sources: T. Shaska, Graded Keller maps and the Jacobian Conjecture, arXiv:2607.20210 (v1 2026-07-22, v2 2026-07-25); the anonymous note at ulam.ai/research/jacobian.pdf (2026-07-20); MathOverflow 513387 (2026-07-20); A. Lou (2026-07-20); and Mayner. PROVENANCE.md §2b gives the claim-by-claim table with dates and theorem numbers. We retain those results for self-containedness and credit priority to those sources.
  • What is actually new here (as of 2026-08-04): (1) the degree-minimality theorem in the weight-(1,−1,−2) equivariant class — no Keller lift of degree ≤ 6 exists at all in the sector containing Alpöge's map, and every degree-≤6 lift elsewhere in the class is an automorphism, by eight msolve unit-ideal certificates over Q reproduced mod 32003; (2) the moment-map identity ν∘Φ = μ and the no-go lemma (s-free C with s-affine A, B ⟹ automorphism, all weights k, all degrees); and (3) a reframing, not a new object: Mayner's Φ identified as an explicit PC_n witness via Adjamagbo–van den Essen, with degrees, momenta, and the quantization identity gr Ψ_F = Φ*. The k = 1, 3 certificates are in progress and are not claimed.
  • Open, and not claimed settled here: JC_2, DC_1 (Zheglov's claimed proof is under review), DC_2, and unconditional minimality of degree 7 among all counterexamples in C^3. Note also that PC_n false for n > 2 is not our result — it follows formally from Mayner's ¬DC_3, and is already asserted in secondary sources.

The claims table

Each row: claim → certificate/artifact → how to re-run. All scripts are pure SymPy and terminate with all asserts passing; the msolve certificates are re-run from their stored input files. The "Origin" column records who first stated the claim publicly — "ours" means we have found no earlier public source; everything else is credited. Verification status and priority are independent: a row marked Mayner or Shaska is still machine-checked here.

# Claim Origin Certificate / artifact Re-run
1 F is Keller: det JF = −2 identically; matches the published form; degrees (7,6,4) Alpöge (2026-07-19) scripts/core_verify.py (PASS lines) python scripts/core_verify.py
2 Rational triple collision F(1,−3/2,13/2) = F(−1,3/2,13/2) = F(0,0,−1/4) = (−1/4,0,0) Alpöge (2026-07-19) scripts/core_verify.py same
3 Master equation (all weights k): det JG = C^k · [C·J(B,A) + k·A·J(B,C) + B·J(C,A)]; example bracket = 2 Shaska, arXiv:2607.20210v2 Thm. 8.3 (2026-07-25); divisorial form v1 Thm. 6.1 (07-22) scripts/core_verify.py same
4 Critical-line restriction, Wronskian = 2; no-go lemma layer identities no-go lemma ours; critical-line contraction in Shaska Prop. 5.1(1) scripts/core_verify.py same
5 Minimal cubic of the cover, trace identity (no u² term: fiber u-values sum to 0) phenomenon in Mayner §4.2 (coord. x) and Shaska Rem. 5.4 (coord. s); u-coordinate version ours scripts/cover_verify.py python scripts/cover_verify.py
6 disc = −4·Δ₁·Δ₂², Δ₁ irreducible ⇒ monodromy full S₃ (cover not Galois) MathOverflow 513387 and Lou (both 2026-07-20); also Mayner §4.4, Shaska Thm. 4.4(2) scripts/cover_verify.py same
7 s ∈ Q(u,p,q) ⇒ F is generically exactly 3:1 ulam.ai Thm. 4.2 / MO 513387 (2026-07-20) scripts/cover_verify.py same
8 Exact image: F misses precisely the punctured curve (4/(27t²), 4/(3t), t), t ≠ 0 ulam.ai Thm. 4.2 (2026-07-20); also Mayner §4.3, Shaska Prop. 5.1 scripts/cover_verify.py (+ preprint §5) same
9 Ψ_F is a well-defined unital endomorphism of W₃: (A) [D_j,F_i] = δ_ij, (B) [D_i,D_j] = 0 (all 9 + 9 identities) Mayner (2026-07-20/21) scripts/weyl_verify.py; expanded operators in scripts/weyl_endomorphism.txt python scripts/weyl_verify.py
10 Cotangent lift Φ(q,p) = (F(q), (JF)^{-T} p) is polynomial, MᵀΩM = Ω exactly, det M = 1 Mayner, REPORT.md §8 (2026-07-21) scripts/dixmier_symplectic_verify.py python scripts/dixmier_symplectic_verify.py
10b Component degrees of Φ are (7,6,4,9,10,12) ours scripts/dixmier_symplectic_verify.py same
11 Φ has a rational triple collision in C⁶ (generically 3:1); Φ* is therefore an injective non-surjective Poisson endomorphism — a PC_n witness non-injectivity of Φ is Mayner's; PC_n false for n > 2 follows formally from Mayner's ¬DC_3 via Adjamagbo–van den Essen and is already asserted in secondary sources; the explicit identification, the momenta, and gr Ψ_F = Φ* are ours scripts/dixmier_symplectic_verify.py same
12 Moment-map preservation: ν∘Φ = μ with μ = xp₁ − yp₂ − 2zp₃ (Hamiltonian C*-equivariance) ours — no occurrence of "moment map" found in the 2026-07/08 literature on this example scripts/dixmier_symplectic_verify.py same
13 Degree minimality in the equivariant class: no Keller lift of degree ≤ 6 in the s-in-C sector; eight unit-ideal certificates over Q ours — cf. Shaska Thm. 10.10 (weaker: A, B degree one in the invariants; general case explicitly open) and Mayner §7 (different ansatz, covering degree) certificates/ms_*_c0.mscertificates/out_*_q.txt (each basis = [1]) msolve -g 2 -f certificates/ms_I-A2c-g1_c0.ms -o out.txt etc. (8 files)
14 Same eight certificates reproduced mod 32003 ours certificates/ms_*_c32003.mscertificates/out_*_p.txt (each basis = [1]) same, on the _c32003 inputs
15 Positive control: the degree-7 example satisfies the identical normalized degree-7 system (pipeline provably contains the counterexample) ours scripts/min_verify.py part d7control (exact scaled substitution, asserted); computational record in certificates/D7CONTROL-NEGATIVE-RESULT.md (the msolve reduced-basis run on certificates/ms_D7control_c32003.ms did not terminate in 600 s — expected for a nonempty variety) python scripts/min_verify.py d7control

The eight emptiness certificates are the six Branch-I systems I-A2c-g1, I-A2c-g2, I-A2L-g1, I-A2L-g2, I-B2-g1, I-B2-g2 and the two full-generality Branch-II systems II-f0, II-f1 (19 and 18 unknowns). Naming: *_c0 = characteristic 0 input, *_c32003 = mod 32003 input; out_*_q.txt = char-0 output, out_*_p.txt = mod-32003 output. A stored output whose reduced Gröbner basis is [1] is a Nullstellensatz certificate that the corresponding system is empty.

Quickstart

python3 -m venv venv
venv/bin/pip install sympy

venv/bin/python scripts/core_verify.py                 # map, det, collisions, master equation, no-go identities
venv/bin/python scripts/cover_verify.py                # cubic, trace, discriminant, S3, 3:1, image locus
venv/bin/python scripts/weyl_verify.py                 # Weyl endomorphism: CCR (A) and (B), writes weyl_endomorphism.txt
venv/bin/python scripts/dixmier_symplectic_verify.py   # symplectomorphism, C^6 collision, moment map
venv/bin/python scripts/min_verify.py ident            # anchor lemma, layer identities, image certificate at (7/3, 4/27)
venv/bin/python scripts/min_verify.py d7control        # degree-7 positive control (exact scaled substitution)
venv/bin/python scripts/min_verify.py kdet             # det JF = -bracket_k upstairs, k = 1,2,3
venv/bin/python scripts/min_verify.py axis             # axis-target uniqueness certificate
venv/bin/python scripts/min_verify.py I                # SymPy GB reproduction, six Branch-I leaves, mod 32003
venv/bin/python scripts/min_verify.py II               # SymPy GB reproduction, two Branch-II leaves, mod 32003

# msolve (https://github.com/algebraic-solving/msolve), v0.10.1 used here:
brew install msolve          # macOS; or build from source
for f in certificates/ms_I-*_c0.ms certificates/ms_II-*_c0.ms;         do msolve -g 2 -f "$f" -o "${f%.ms}.out"; done   # 8 unit ideals over Q
for f in certificates/ms_I-*_c32003.ms certificates/ms_II-*_c32003.ms; do msolve -g 2 -f "$f" -o "${f%.ms}.out"; done   # mod-32003 reproductions

Every Python script must end with its PASS lines and no assertion failures. Every ms_I-*/ms_II-* msolve run must output the reduced basis [1], matching the stored out_*_q.txt (char 0) and out_*_p.txt (mod 32003). Do not loop ms_D7control_c32003.ms in: that reduced-basis run does not terminate in reasonable time (see certificates/D7CONTROL-NEGATIVE-RESULT.md); the asserted positive control is scripts/min_verify.py d7control.


Part B — certified quantum code distances

Replayable minimum-distance certificates for stabilizer codes, with no solver and no proof assistant in the trusted base: the bivariate-bicycle family through n = 288. Paper: paper/preprint-qec-distances.pdf.

This is a verification contribution, not a discovery. The distance values are largely known and are credited below. What did not exist is a standalone artifact anyone can replay without a SAT solver and without a proof assistant.

Lead with the audit

The corpus was re-checked by a separate agent instance with no access to the generating pipeline and no shared code. Its report is INDEPENDENT-VERIFICATION.md.

Certificate checks run 47
Passed / failed 47 / 0
LRAT bytes replayed 5,459,315,046 (5.08 GiB), in pure Python — no compiled code in the trusted base
Peak resident set, worst case 79 MB (the 2.94 GB proof)
IBM BB parity-check matrices vs. arXiv:2308.07915 byte-identical, all five codes
manifest.json SHA-256 entries re-hashed 182 / 182 match
Negative controls 11 / 11 correctly rejected
Brute-force cross-checks (codes small enough) 6 / 6 agree

Results

code n, k certified lower-bound proof in this repo?
Steane [[7,1,3]] 7, 1 d = 3 699 B each sector yes
five-qubit [[5,1,3]] (non-CSS) 5, 1 d = 3 1,514 B yes
rotated surface d=3 / 5 / 7 9 / 25 / 49, 1 d = 3 / 5 / 7 518 B – 75 kB yes
Golay [[23,1,7]] 23, 1 d = 7 226 kB each yes
IBM BB [[72,12,6]] 72, 12 d = 6 1.5 / 1.9 MB yes
IBM BB [[90,8,10]] 90, 8 d = 10 128 / 151 MB yes (gzipped)
IBM BB [[108,8,10]] 108, 8 d = 10 72 / 82 MB yes (gzipped)
IBM gross [[144,12,12]] 144, 12 d = 12 124 MB symmetry-broken; 868 / 672 MB symmetry-FREE symmetry-broken yes; symmetry-free regenerable
IBM BB [[288,12,18]] 288, 12 14 ≤ d ≤ 18 2.94 GB (K = 13) K=9 rung yes; K=11, K=13 regenerable

Four proofs (79 MB–646 MB compressed) are too large for git. qec-certificates/REGENERATE.md gives the exact CaDiCaL command, expected byte count, and expected SHA-256 for each. Everything else ships, so a reader with nothing but CPython can still replay a certified d = 12 for the gross code.

Credit — read this first

  • The codes and the distances are IBM's. The bivariate-bicycle family, including the gross code, is Bravyi, Cross, Gambetta, Maslov, Rall and Yoder, Nature 627 (2024) 778 / arXiv:2308.07915. Their distances were computed there by the MIP method of Landahl, Anderson and Rice (arXiv:1108.5738). The gross-code value d = 12 was confirmed exactly, at MIP gap 0, by Cruz-Benito, Cross, Kremer and Faro (IBM, arXiv:2606.02418, 1 Jun 2026). We claim no distance value.
  • d_X = d_Z for BB codes is Bravyi et al.'s lemma, from their supplemental material. Only the explicit permutation, packaged as a ~15 ms checkable certificate, is ours.
  • Machine-checked quantum distance proofs are LEAN-QEC's (arXiv:2605.16523). Their paper dispatches the gross code to cvc5 outside the Lean kernel and calls kernel replay "the next concrete engineering target" — but their repository has moved past their paper: commit c73827d (2026-07-10) records a full [[144,12,12]] verification via bv_decide in about 30 minutes, including kernel replay. We claim no priority for a machine-checked gross-code distance. What differs, at that commit: their BB144.lean carries two sorrys (BB144_X_ker_rank L69, BB144_Z_ker_rank L72) that BB144_dist_12 routes through; three lemmas use native_decide, which their own paper notes extends the trusted base with Lean's compiler; no LRAT artifact is committed for BB144; and their encoding is symmetry-broken only. Ours has no admitted lemmas, ships the artifacts, includes symmetry-free proofs, and needs no proof assistant. Their kernel-checked ladder should not be described as reaching n = 108 either: BB108.lean carries sorry at L120 and L132 with --bv_decide commented out.
  • [[288,12,18]], stated correctly. Bravyi et al. assert d = 18 exactly, by ILP, without shipping a checkable artifact — their Table 3 lists it with no "≤", unlike [[360,12,≤24]]. Our interval [14,18] is not new information about the value. What is defensible, and all we claim: Chen, Jafari and Lai (arXiv:2606.12445) report d ≥ 11 solver-asserted with no proof artifact in their repository; we certify d_X ≥ 14 with a 2.94 GB LRAT that replays independently — improving the strongest quantity previously published as a lower bound, and the only machine-checkable one on record for this code.
  • Also prior art, cited at point of use: QDistRnd (JOSS 2022, upper bounds only, "no performance guarantee"); Stim's search_for_undetectable_logical_errors (documented verbatim as "THIS IS A HEURISTIC METHOD"); the Webster–Jacob–Higgott survey (arXiv:2603.22532); PBLean (arXiv:2602.08692); Heule's drat-trim/LRAT; Biere's CaDiCaL; Sinz's cardinality encoding; Tseitin's gate encoding.

The dated adversarial sweep behind these statements is SWEEP-RECORD-QEC-2026-08-04.md. It broke two claims of an earlier draft; both were withdrawn.

Quickstart (Part B)

No virtual environment, no packages, no compiled binary — the checkers import only the Python standard library.

# small, instant
python3 qec-scripts/check_witness.py  qec-certificates/steane/witness_X.json
python3 qec-scripts/check_lower.py    qec-certificates/steane/lower_X_K2.json
python3 qec-scripts/check_lower.py    qec-certificates/golay/lower_X_K6.json

# the gross code, d = 12, from artifacts in this repository
gunzip qec-certificates/bb144/*.lrat.gz
python3 qec-scripts/check_witness.py  qec-certificates/bb144/witness_X.json     # d_X <= 12
python3 qec-scripts/check_lower.py    qec-certificates/bb144/lower_X_K11_sym.json  # d_X >= 12
python3 qec-scripts/check_duality.py  qec-certificates/bb144/duality.json       # d_X = d_Z

# n = 288
gunzip qec-certificates/bb288/*.lrat.gz
python3 qec-scripts/check_lower.py    qec-certificates/bb288/lower_X_K9_sym.json   # d_X >= 10
python3 qec-scripts/check_duality.py  qec-certificates/bb288/duality.json

# integrity: re-hash every artifact against the audited manifest
python3 qec-scripts/verify_manifest.py

verify_manifest.py reports 172 match, 0 mismatch, 10 absent on a fresh clone: the ten "absent" are the six proofs that ship gzipped (decompress them and they match, since the manifest hashes the uncompressed bytes) and the four that are too large for git. After gunzip qec-certificates/*/*.lrat.gz the count is 178 match, 4 absent.

Note that qec-scripts/manifest.py is the pipeline's manifest generator, not a verifier: it expects a certificates/ directory beside itself and it overwrites manifest.json. It is included because it is part of the pipeline, and the pipeline is explicitly not trusted.

check_lower.py never reads the shipped .cnf: it regenerates the CNF from the raw parity-check matrices and replays the LRAT against its own clause list. Corrupting a shipped .cnf changes nothing, and the audit confirmed that by doing it.

Known gaps in Part B

  1. bb288/duality.json was generated after the audit closed — it passes check_duality.py and its permutation was independently re-verified, but it is outside the 47 audited checks and absent from manifest.json.
  2. manifest.json covers the full audited corpus including the four large proofs this repository does not carry.
  3. The shipped check_lower.py is 481 lines; the auditor read 419. The difference is an optional totalizer cardinality encoding that no certificate in this release selects. The whole corpus was re-run against the shipped 481-line checker before release — 20 witnesses, 5 duality certificates, 23 lower-bound certificates, 48 in all, every one accepted, including a second pure-Python replay of the 2.94 GB proof.
  4. run_all.sh expects a tools-drat-trim/lrat-check binary that is not vendored. The pure-Python path needs nothing but CPython and is what every number in the paper reports.

Three further defects the audit found are reported verbatim in §8 of the paper, including a latent soundness hole in a checker branch that no shipped certificate exercises. A paper about trusted bases that suppresses its own audit findings is not one.


Layout

README.md            this file
PROVENANCE.md        timeline, what is new, how to verify, what is not claimed (§§1-4 Part A, §5 Part B)
CITATION.cff         citation metadata
.zenodo.json         Zenodo deposit metadata
LICENSE-CODE         Apache-2.0 (code and machine-readable certificates)
LICENSE-DOCS         CC-BY-4.0 (prose and paper)
paper/               both preprints (LaTeX + PDF + readable Markdown mirror)

  -- Part A: Alpoge Keller --
certificates/        17 msolve input files (ms_*.ms) + stored outputs (out_*.txt) + D7 control record
scripts/             the verification scripts (incl. min_verify.py) + expanded Weyl operators
schema/              certificate-schema (pending; see schema/PENDING.md)
checker/             independent certificate checker (pending; see checker/PENDING.md)

  -- Part B: quantum code distances --
qec-certificates/    the certificate corpus, by code; manifest.json; REGENERATE.md
qec-scripts/         the three checkers (check_witness / check_lower / check_duality),
                     verify_manifest.py, and the (untrusted) generating pipeline
INDEPENDENT-VERIFICATION.md   the audit: 47/47 checks, 5.08 GiB replayed in pure Python
SWEEP-RECORD-QEC-2026-08-04.md   the dated adversarial prior-art sweep

For Part A, the reduction library and the system generators are intentionally not part of this repository; the published claims are the certificates themselves plus the verification scripts, which are self-contained. For Part B, the generating pipeline is included (qec-scripts/certify.py, qec_lib.py, run_all.sh) precisely because it is not trusted: it can be deleted and every certificate still verifies.

Licensing

Dual license by content type:

  • Code and machine-readable certificate files — everything under scripts/, certificates/, schema/, checker/, qec-scripts/, qec-certificates/ — are licensed under the Apache License 2.0 (LICENSE-CODE).
  • Documentation and the paperpaper/, README.md, PROVENANCE.md, and all other prose — are licensed under CC BY 4.0 (LICENSE-DOCS).

Citing

See CITATION.cff. A DOI will be minted on the first Zenodo deposit; until then, external timestamps for this repository's claims begin at the first public push — not at local file dates (see PROVENANCE.md §3).

Contact

Daniel Kirtchakov — daniel@halfounce.io

About

Certificate-first verification: degree minimality in the equivariant class of the Alpöge Keller map — every claim ships with a re-runnable certificate

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages