cert-machine · six claims, six verifiers written from scratch

We checked the AI's homework

Six mathematical results produced with help from frontier AI systems — a counterexample to Maxwell's point-charge bound, a new lower bound for the Korenblum constant, an Erdős problem from 1958, and three more. Each one was re-verified here in exact arithmetic, from the manuscript, by verifiers that never ran a line of the authors' code. This is what survived.

tl;dr
  • The finding. 5 of 6 held; 1 came back PARTIAL; 0 were refuted. But the number that matters is the other one: in 6 of 6 cases the computational fragment certified and the analytic core did not — the identities, enclosures and exhibits are machine-checkable and they all held, while the surrounding argument is prose in every single one. That split is the finding.
  • The mechanism. Each verifier was written against the manuscript, never derived from and never executing the author's scripts — no Arb, no SymPy, no Mathematica, no code from anybody's archive was run. (In one lane, Mathieu, the author's Python was read to cross-check which identities the TeX meant; it was not executed, and that is stated on the lane.) Exact rationals where the claim is algebraic, outward-rounded intervals and the Krawczyk operator where it is analytic. Each carries mutation controls: deliberate corruptions of the claim that the verifier must reject. A verifier that cannot fail is not evidence.
  • Check it. node instruments/laneaudit/audit.js — all six, 6.0 s end to end on a laptop. 92 individually named checks and 21 mutation controls, every control rejected. This page is built from that run and refuses if any of it goes red.
claims re-verified
6
each written from the manuscript; no author code was executed in any lane
held
5 confirmed
within a scope this page states for each one, never wider
partial
1
Ran–Teng: every machine-checkable fragment holds; the analytic core is prose
refuted
0
no claim in this set was contradicted by our arithmetic
analytic cores audited
0 of 6
the honest ceiling of the whole exercise, and the reason this page exists
mutation controls
21
deliberate corruptions the verifiers must reject — all 21 rejected
named checks
92
counted by one rule across all six, not taken from six verifiers that disagree about what a check is
§1 · the board

Six claims, and exactly how far the checking reached

claimAI systemverdictscope of that verdictnamed checksits own totalcontrols
Maxwell's point-charge boundOpenAI GPT-5.6 Sol (idea)CONFIRMEDat ε = 1/6 only17203
The Korenblum constantOpenAI Codex (GPT-5.6 Sol), disclosed in the paperCONFIRMEDthe numerical criterion154
Erdős Problem #1038GPT-5.5 Pro (initial main argument in a solver–verifier framework, then revised by the human authors)CONFIRMEDthe computational fragment22264
Ran–Teng Conjecture 20GPT-5.2 ThinkingPARTIALmachine-checkable fragment only38435
The Mathieu property for Lie groupsnot recorded at sourceCONFIRMEDsupporting identities only1263
The rank-two Poisson conjecturenot recorded at sourceCONFIRMEDthe explicit counterexample2

The six verifiers do not agree about what a check is, and the two columns say so rather than hiding it. Maxwell, Ran–Teng and Erdős #1038 fold their mutation controls into their own printed total (17+3=20, 38+5=43, 22+4=26); Mathieu counts its controls separately; Korenblum and Poisson print no total at all. So the first column is computed here by ONE rule applied to all six — a named check row that is not a mutation control — and the second column is whatever each verifier says about itself. Where a verifier names no rows, or states no total, the cell reads instead of a number we invented. Click any claim for its full ledger.

Scope is the whole product here. Every verdict above is true inside the phrase beside it and false outside it. "CONFIRMED at ε = 1/6" is not "CONFIRMED"; "supporting identities only" is not "theorem proved". The claims themselves are stated with more confidence than that at source, which is precisely why an independent scope line is worth having.

0 10 20 30 40 Ran–Teng Conjecture 20 38 Erdős Problem #1038 22 Maxwell's point-charge bound 17 The Korenblum constant 15 individually named checks the verifier printed at build time (hover for controls and runtime)
Only four of the six print named check rows, so only four are plotted — Mathieu reports 126 checks as a summary and Poisson reports its identities as a set, and a bar drawn from a number that names nothing would be a bar drawn from nothing. Ran–Teng, in the contrasting colour, is the one PARTIAL, and it carries the most named checks in the set: depth of checking is not depth of proof.
§2 · the finding

The fragment always certifies. The core never does.

the computational fragment the analytic core Maxwell's point-charge bound CONFIRMED The Korenblum constant CONFIRMED Erdős Problem #1038 CONFIRMED Ran–Teng Conjecture 20 PARTIAL The Mathieu property for Lie groups CONFIRMED The rank-two Poisson conjecture CONFIRMED certified here, in exact arithmetic not audited — the argument is prose
Every row is one claim. Left half: the part a machine can decide, and did. Right half, hatched: the part that carries the theorem. Six for six, with no exception in either column.

This is not a complaint about the authors and it is not a complaint about the AI. It is a measurement of where the checkable surface of a modern mathematical claim actually stops. The exhibits — an eight-atom measure, twenty-four critical points, a quartet of polynomials, thirty decimals of an extremal constant — are all real, and every one of them survived an independent implementation that was trying to break it. The arguments that turn those exhibits into theorems are English prose with mathematical content, and no verifier in this repository or anywhere else can read them.

The one lane where that gap closes is Poisson, and it closes for a structural reason worth naming: a counterexample is an existence claim. Certifying the exhibit IS certifying the claim. Every other lane in this set asks a machine to confirm a universal statement from a finite computation, which it cannot do — so the verdict has to carry a scope line, and it does.

§3 · claim by claim

Pick a claim: what was checked, and what was not

Maxwell's point-charge bound

CONFIRMED at ε = 1/6 only

The claim. Five positive point charges in ℝ³ whose potential has at least 24 nondegenerate critical points — exceeding the conjectured bound of 16. Maxwell conjectured that n point charges produce at most (n−1)² critical points of their potential.

Credited system. OpenAI GPT-5.6 Sol (idea); written up by Arathoon, Ball and Kvalheim
Verification at source. Three-author verification; Mathematica and Maple checks — floating point, no code artifact, no certificate, and no concrete ε.

what we certified

At the fixed rational ε = 1/6 with qε = 859/248832 exactly, 24 pairwise-disjoint boxes, each certifying existence and local uniqueness of a critical point by the Krawczyk operator and nondegeneracy by an interval Hessian determinant that excludes zero. 24 > 16 is then exact integer arithmetic.

what we did NOT audit

The theorem’s asymptotic statement — "for all sufficiently small ε" — and every proof in the paper. One ε is certified, not a range.

The Korenblum constant

CONFIRMED the numerical criterion

The claim. c₂ ≥ 0.4263, via a moment-duality criterion and an explicit rational eight-atom measure (arXiv:2607.17748). The Korenblum constant c₂ had a best published lower bound of 0.3554.

Credited system. OpenAI Codex (GPT-5.6 Sol), disclosed in the paper; author proof by Frank Wikström
Verification at source. An Arb interval certificate shipped with the paper, plus a Zenodo archive.

what we certified

Every inequality of the paper’s moment-duality criterion for its explicit measure, under our own outward-rounded interval arithmetic — no Arb, no python-flint, and no code from the Zenodo archive executed. The A-family over k = 1..2183, the B-family over k = 1..1299, both tails by exact induction.

what we did NOT audit

The duality argument itself — the paper’s Lemma 3.1 and Wang’s Proposition 2.1. The bound follows from what we certified only if that argument holds.

Erdős Problem #1038

CONFIRMED the computational fragment

The claim. The infimum is exactly D = 1.834430475762661711090753635125…, in a July 2026 manuscript of Darvas, Peng and Tao. Among monic polynomials with all roots in [−1,1], what is the infimum of the measure of the set where |f| < 1? Open since Erdős–Herzog–Piranian, 1958.

Credited system. GPT-5.5 Pro (initial main argument in a solver–verifier framework, then revised by the human authors)
Verification at source. Author-checked manuscript shipping its own three Arb certificates and two SymPy verifiers; the official Erdős-problems record was still open.

what we certified

The extremal triple exists, is unique inside a certified box, lies in the admissible region, and every printed decimal of the extremal constants is correct — including all 30 claimed decimals of D.

what we did NOT audit

The proof that this triple is extremal. We certified the number the manuscript names; we did not certify that no polynomial does better.

Ran–Teng Conjecture 20

PARTIAL machine-checkable fragment only

The claim. Resolved, in a preprint. Source status: "Human-checked mathematical proof; no formal proof assistant artifact located" (24 Feb 2026). The exact nonreal spectral region for a family of structured matrices — a conjecture in matrix analysis.

Credited system. GPT-5.2 Thinking
Verification at source. Human-checked prose. No artifact of any kind was located.

what we certified

Every load-bearing polynomial identity of the proof, exactly over ℚ. Both boundary-attainment families re-proved exactly, including an exact re-proof that every nonreal eigenvalue of the A_L family lies ON the conjectured curve. Certified eigenvalue enclosures over 272 exact matrices found zero counterexamples to the necessity conditions, and attainment is Krawczyk-certified at 10 interior points.

what we did NOT audit

The analytic core — the argument parametrization, the convexity/Jensen/Karamata optimization that makes necessity hold for ALL parameters, and the branch bookkeeping. It is prose. This is why the verdict is PARTIAL and not CONFIRMED.

The Mathieu property for Lie groups

CONFIRMED supporting identities only

The claim. Exactly the tori — a classification. For which compact connected Lie groups G is the kernel of Haar integration on R(G) a Mathieu–Zhao space?

Credited system. not recorded at source
Verification at source. An author SymPy script, which we read but did not execute.

what we certified

Every supporting identity the classification rests on, exactly over ℚ with BigInt rational Laurent arithmetic: the weighted witness moments, the explicit abelian pair and its printed four-term expansion, the matrix-entry representatives and their invariance. Disclosure specific to this lane: the author’s Python was READ to cross-check which identities the TeX intended — its blocks match the manuscript and are the ones re-verified here — but it was never executed, and no value on this page came from it.

what we did NOT audit

The classification theorem itself. Confirming the identities a proof uses is not confirming the proof, and this lane is the clearest case of that distinction in the set.

The rank-two Poisson conjecture

CONFIRMED the explicit counterexample

The claim. An explicit counterexample: polynomials R, T, D, S in ℚ[x,q,p,z] with prescribed bracket relations. A conjecture about rank-two Poisson structures.

Credited system. not recorded at source
Verification at source. The author’s own scripts, which we neither derived from nor executed.

what we certified

All six bracket relations ({D,R} = 1, {S,T} = 1 and the four vanishing ones) hold identically and exactly; the 4×4 symbolic Jacobian determinant is identically 1; and the three claimed points each map exactly to their stated image. A counterexample is an existence claim, so certifying the exhibit IS certifying the claim — the only lane in this set where that is true.

what we did NOT audit

The author’s surrounding prose beyond the extracted definitions. Nothing load-bearing: the counterexample stands or falls on the exhibit, and the exhibit holds.

§4 · what moved

Two of the six move a published number

0% 20% 40% 60% Maxwell's point-charge bound 16 → 24 (+50.0%) The Korenblum constant 0.3554 → 0.4263 (+19.9%) relative increase over the number the claim replaces
The two rows are in different units, so the axis is the one quantity they share: how far each claim moves the number it replaces. The absolute values sit on the bars.

For Maxwell the move is the whole point: 24 > 16 is exact integer arithmetic once the twenty-four boxes are certified, so at ε = 1/6 the conjectured bound is false, and that much is now independently established rather than asserted from floating-point computer algebra. For Korenblum the move is conditional: our arithmetic puts the criterion beyond doubt and the bound follows only if the paper's duality lemma holds, which we did not audit.

The other four claims do not move a number at all — they settle a question that had no previous numerical answer, or exhibit an object whose existence was open. There is nothing to plot for those, and inventing an axis for them would be the kind of chart this repository does not draw.

§5 · re-run it

Six seconds, one command, no dependencies

The verifiers are plain Node with no packages. Clone the repository and run them; there is nothing to install and nothing to trust from us except the arithmetic, which is the point.

git clone https://github.com/carlostoledo1891/cert-machine
cd cert-machine
node instruments/laneaudit/audit.js

Each verifier can also be run alone, in its own directory, and prints its own report — the numbers on this page are parsed out of exactly those reports at build time:

node legacy/research/challenges/lane/laneb-maxwell/verify.js
node legacy/research/challenges/lane/laneb-korenblum/verify.js
node legacy/research/challenges/lane/laneb-lemniscate/verify.js
node legacy/research/challenges/lane/laneb-ranteng/verify.js
node legacy/research/challenges/lane/laneb-mathieu/verify.js
node legacy/research/challenges/lane/laneb-poisson/verify.js

To confirm the verifiers can actually fail, corrupt one and watch it refuse. That is what the 21 mutation controls do automatically on every run: perturb a coefficient, flip a bracket sign, move mass between atoms, replace a candidate with a duplicate. All 21 were rejected in the run that produced this page.

what this page does NOT claim

We did not prove any of these theorems and we did not refute any of them. No analytic argument in the set was audited, because none of them is machine-checkable at any budget we have. What we did is narrower: take six results a frontier model helped produce, write independent verifiers from the manuscripts, and report exactly which parts survive exact arithmetic and which parts were never in reach. The verdicts are ours; the theorems remain the authors'. Several of these claims were already checked by their authors, and in two cases by an Arb certificate the authors shipped — an independent reimplementation is worth something anyway, but it is not a first verification and this page never says it is.