Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

OpenAI's Yang-Mills proof is public in Lean

Written November 11, 2026

The committee of five physicists appointed to check OpenAI's proof has released the full 1,147-page Lean formalization with a 35-page assessment affirming that pure Yang-Mills theory exists. The mass gap — the other half of the Millennium Prize problem — remains open, and the reviewers say they do not yet fully understand the proof.

What was released

On the night of November 10, the Yang-Mills Working Group — David J. Gross, who chairs it, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov — published OpenAI's full Lean code in the public repository yang-mills-group/OpenAI-YM-20261028, alongside their assessment. The timeline behind the release: OpenAI says its model produced the natural-language existence proof on September 9 and completed the Lean formalization on October 16, the same day it asked Gross to take custody of the proof and shepherd its release. OpenAI announced the result on October 26 — the existence half only — and held the artifacts back while the committee reviewed them. The committee's statement had promised the release for November 10, and it arrived on schedule.

The assessment judges that the construction proves the existence of pure Yang-Mills theory according to the problem statements standard in physics and mathematics. It rests its judgment on a stronger combined target that binds the lattice family, the limiting correlations, and the reconstructed theory together, rather than on the original joint-existence target alone, which omitted some lattice-construction links established elsewhere in the proof.

What stays open

The Yang-Mills Millennium Prize problem has two parts: prove the theory exists, and prove a mass gap emerges from its equations. This release addresses the first. On the mass gap, the assessment is explicit: the construction supplies no evidence for or against one. Confinement is likewise unresolved, and the construction does not prove the infinite-volume limit is unique. The repository's own summary says the same in plain terms: a positive mass gap, general exclusion of continuous-mass generalized free fields, and all-orders matching are outside its scope.

That is why this is not a Millennium Prize solution. A complete prize claim would need the missing mass-gap proof, a qualifying publication, at least two years, and general acceptance by the mathematics community. Waiting alone cannot supply the missing theorem.

What Lean checking does and does not do

Lean is a proof assistant: its kernel checks whether a formal proof establishes a precisely encoded proposition from the accepted assumptions. The repository reports that the pinned Lean kernel replayed 100,168 project declarations for the 17 advertised results and found only the standard axioms of propositional extensionality, classical choice, and quotient soundness. The standalone Comparator executable was not run. That is a real check, and it answers a narrower question than the one a physicist asks first: whether the proof term establishes the proposition under those assumptions. It cannot decide whether the proposition captures the intended Yang-Mills construction — reviewers have to inspect the definitions, their connections, and what the theorem leaves out.

The reviewers are candid about where that inspection stands. They used OpenAI's GPT-6 Astra model to help digest the Lean code, and they write that they do not yet fully understand the proof. The group reports no current connection to the Clay Mathematics Institute. With the proof and its Lean code now public, the wider community can use its own agents to reverse-engineer and parse it — which is the point of releasing it.

The argument about how it was released

The release method is part of the story. OpenAI held the artifacts back and handed them to a committee it asked Gross to lead, rather than publishing them openly at announcement. The Advisory Group on Mathematics and Artificial Intelligence, an independent panel hosted at the Institute for Advanced Study, had recommended immediate release of the natural-language and Lean outputs, on the grounds that understanding should develop in the open with everyone given the same access. As Scientific American reports, the shepherding approach contradicts that recommendation and has drawn criticism from mathematicians who see it as needless gatekeeping. That is a process objection, not a technical refutation of the proof — and it is worth carrying separately from the mathematics.

A parallel thread runs alongside the committee's release. A separate Lean formalization by Victor Logvinovich, updated November 8, argues that the Clay conjunction on non-compact R⁴ cannot hold — strict Osterwalder–Schrader/Wightman axioms and a positive mass gap are mutually exclusive in the formal class the author defines. That is an unresolved thread from a different author, not a refutation of the committee's assessment, and it should be read as one more open question in a field that now has the code to argue from.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · AlphaProof Code verified Signal in Lean