# Yang-Mills gap audit: the OpenAI existence claim against the Clay criteria

Date of audit: 2026-11-20. Status of this document: background reading for the Millennium Lab mass gap hunt, not a verdict on the Clay problem and not a substitute for reading the sources. No solution claim is made anywhere in this file.

## What was read

1. OpenAI announcement, October 26, 2026: "Working with academic communities to build scientific understanding". Read 2026-11-20 at https://openai.com/index/building-scientific-understanding-with-academic-communities/
2. The Lean formalization submitted October 28, 2026, in the public repository yang-mills-group/OpenAI-YM-20261028. Read 2026-11-20 at https://github.com/yang-mills-group/OpenAI-YM-20261028 (repository listing and README description only; the Lean code itself was not read line by line for this audit).
3. The independent assessment, November 10, 2026: Douglas, Jaffe, Nissim, and Rajasekaran, "Validation of the OpenAI Yang-Mills construction", approved by the Yang-Mills group, with consultative review from Gross, Klebanov, and Nekrasov. Read 2026-11-20 at https://raw.githubusercontent.com/yang-mills-group/YM/main/docs/assessment-1.pdf (summary, sections 1 through 6, section 8 open questions, and sections 9 and 10 on validation).
4. The critique, November 2026: Jenny Lorraine Nielsen, "The OpenAI Yang-Mills Formalization Does Not Establish Existence", Center for Topological Physics, Lawrence, Kansas. Read 2026-11-20 at https://philarchive.org/archive/NIETOA-3 (abstract and sections 1 through 11).
5. The Clay rules as summarized in the Millennium Prize Problems reference: the Clay Institute only considers proposed solutions after at least two years from publication, and requires solutions in the form of peer reviewed publications that are then further scrutinized. Found by search 2026-11-20 at https://en.wikipedia.org/w/index.php?title=Millennium_Prize_Problems&oldid=1375752114

What was not read: the 1,147 page informal manuscript, which has not been released, and the full Lean source. The assessment states explicitly that it does not draw on the informal manuscript, and that the group does not claim to fully understand the proof yet. This audit therefore judges the assessment and the critique against the Clay criteria, not the proof itself.

## What happened, in order

On September 9, 2026, OpenAI says its models produced a natural language proof of Yang-Mills existence, and on September 18 a proof for quantum chromodynamics. The Lean formalization of the Yang-Mills result was finished October 16, and OpenAI asked David Gross that day to take custody of the proof. On October 26 OpenAI announced three results, including the existence proof for quantum chromodynamics and pure Yang-Mills theory, and described the manuscripts as raw research data to be handed to outside experts. The Lean code was submitted to the Yang-Mills group on October 28. On November 9 the group issued a statement saying the result proves the existence of pure Yang-Mills theory under the standard problem statements. On November 10 the group published the Lean code and a 35 page assessment. The mass gap was explicitly not claimed at any point. A second OpenAI submission dated November 4 is listed as work in progress and is unassessed.

## The Clay criteria

The Jaffe-Witten statement of the problem asks: prove that for any compact simple gauge group G, a non-trivial quantum Yang-Mills theory exists on R4 and has a mass gap Delta greater than zero. Following Nielsen's breakdown, which tracks the Jaffe-Witten text, existence itself has five requirements. E1: the Wightman axioms W0 through W3, including the spectrum condition and uniqueness of the vacuum, field axioms, Poincare covariance, and spacelike locality. E2: non-triviality, meaning the theory cannot be a free or generalized free field theory, linked to an operator product expansion with the singularities predicted by asymptotic freedom, a stress-energy tensor, and agreement with perturbative renormalization at short distances. E3: Yang-Mills identity, meaning the constructed object must be the quantization of the classical Yang-Mills Lagrangian, not merely some quantum field theory that satisfies the Wightman axioms and starts from a lattice. E4: the theory must be defined on R4, so the infinite-volume limit must be taken and must yield a unique, well defined theory. E5: the construction must work for any compact simple gauge group. On top of all five sits the mass gap: a positive lower bound Delta on the mass spectrum above the vacuum.

## What the assessment says is met

The assessment judges that the submission proves a relativistic quantum field theory satisfying the Wightman axioms can be constructed as a scaling limit of four-dimensional Wilson lattice gauge theory for any simple compact gauge group. That covers E5, the universality over gauge groups, and the core of E1, the Wightman axioms, for the component theories produced by the phase decomposition. The construction starts from the Wilson lattice definition, associates renormalized lattice observables to gauge-invariant polynomials in the curvature and its covariant derivatives, proves convergence of their correlation functions at fixed volume, takes a subsequential infinite-volume limit, decomposes the mixed theory into component theories that satisfy the Osterwalder-Schrader axioms including clustering, and applies a reconstructed Osterwalder-Schrader theorem to obtain Wightman theories with vacuum uniqueness, common domain, Poincare covariance, and local commutativity for each component. A conserved stress tensor is constructed, a nonzero trace with a short-distance law is proved, and an operator product expansion with error bounds is provided. The formal check is strong: a clean rebuild with Lean 4.32.0 rebuilt 9,470 project modules in about 184 minutes with zero errors, the inventory found no project axiom declarations and no direct reference to the placeholder axiom sorryAx in the solution, the principal proofs use only the permitted foundational axioms propext, Classical.choice, and Quot.sound, and the Comparator passed on October 30 for the public theorem and on November 3 for the group's stronger combined theorem. The assessment is candid that the mixed theory itself is not proven to have a unique vacuum, and that uniqueness across volume subsequences is unresolved.

## What is disputed

Nielsen argues that the formalization fails existence in the Clay sense for seven independent reasons, and the assessment itself concedes the central facts behind the strongest three. First, convergence is subsequential only. The theorem states that for every volume sequence there exists a subsequence and a hierarchy with the stated properties, with no claim that different subsequences give the same theory. The assessment agrees: the infinite-volume limit is only taken subsequentially, it remains unresolved whether the limiting correlations depend on the chosen subsequence, and its own open questions list establishing convergence and uniqueness of the infinite-volume limit as unfinished work. This is the E4 dispute. Second, the non-triviality test does not exclude generalized free fields. The formalization's trace test excludes only finite Wick polynomial profiles, while a generalized free field has a continuous spectral density and passes that test. The assessment's own text says the construction provides no evidence for or against a mass gap, and the submission README places general exclusion of continuous-mass generalized free fields outside its scope. This is the E2 dispute. Third, the contact terms are uncontrolled. Correlation functions are established only for separated points, and the assessment notes that uniqueness over other contact prescriptions and raw cutoff convergence on collision diagonals are outside the claim, and that the diagonal-flat Osterwalder-Schrader comparison cannot establish uniqueness of the contact extensions. This bears on E1 and E3, since Wightman fields must be operator-valued distributions defined on all test functions. Fourth, the identification of the constructed object as Yang-Mills theory has not been verified by a human mathematician. The assessment offers evidence rather than proof on this point: lattice origin, local fields, normalized stress tensor, leading short-distance laws, and operator product expansion provide substantial evidence that the construction describes Yang-Mills theory, while universality, infrared uniqueness, and full perturbative agreement remain open. Nielsen adds three further objections about the formalization itself: the gauge group bridge and the Wilson measure definition have structural subtleties, the plaquette is implemented as a one-dimensional fold over a list of links which in her reading does not encode the two-dimensional curvature geometry, about 60 percent of the declarations are not on the dependency path of any advertised result, and the theorem is universe polymorphic with the exclusion instantiated at level zero. These last points are about definitional adequacy and human reviewability, not about the Lean kernel check, which both sides accept as passed.

## What is not met

The mass gap is not met. Nobody claims otherwise. The assessment states it plainly: the construction provides no evidence for or against a mass gap in the theory, and its open questions list establishing a uniform mass gap, giving exponential decay of correlations and a spectral gap in the Osterwalder-Schrader Hamiltonian, as unfinished work. The Wilson loop area law is not established, no continuum probability measure is constructed, and no single-particle states or energy-momentum spectrum are analyzed. The human readability gap is also open: the 1,147 page manuscript is unreleased, the group says it does not claim to fully understand the proof yet, and the assessment rests on the Lean code alone. The Clay standard requires a proof that has received general acceptance in the mathematics community, and that acceptance does not exist yet for this submission.

## Could the Clay two-year clock start

No. Two independent reasons block it. First, the Clay rules require a peer reviewed publication, and at least two years must pass from publication before the Institute considers a proposed solution. What exists today is a Lean submission in a public GitHub repository and a 35 page assessment document from a working group that states it has no present connection to the Clay Mathematics Institute. That is not a peer reviewed publication of a solution. Second, even if a publication existed, the submission addresses only the existence half. The Clay problem is a single problem with two parts, existence and mass gap, and the mass gap half has no proof and no claimed proof. The clock cannot start on a partial result for a problem that requires both halves.

## What would have to change for the mass gap half

A mass gap proof would have to establish a uniform spectral gap: exponential decay of the relevant correlations and a gap in the spectrum of the Osterwalder-Schrader Hamiltonian above the vacuum, with a positive Delta that does not collapse as the volume grows. That is the assessment's own open question number three, and it is the hard half of the problem. The surrounding gaps would also have to close for the full Clay problem: full-sequence convergence to a unique infinite-volume limit with a unique vacuum, a continuum probability measure, the Wilson loop area law, a single-particle spectrum, and a human readable exposition that the mathematics community can evaluate and accept. Until those exist, the honest status is the one the assessment itself states: a Wightman quantum field theory constructed as a scaling limit of Wilson lattice gauge theory, with the mass gap open and the identification as Yang-Mills theory supported by evidence rather than closed by proof.
