OpenAI Yang-Mills Proof Verified: What the Expert Panel Actually Confirmed
Key Takeaways
- A five-member committee chaired by Nobel laureate David Gross has verified OpenAI's AI-generated proof of the existence of quantum Yang-Mills theory, and published the Lean code and its assessment on 10 November 2026.
- The proof establishes the first half of the Yang-Mills Millennium Prize problem — the existence part — under the problem statements standard in physics and mathematics.
- It does not prove the mass gap, the harder second half, so the Clay Mathematics Institute's $1 million prize is not won.
- The committee's own account says it used OpenAI's Astra model to help examine the code and that it does "not claim to fully understand the proof yet".
- OpenAI submitted the proof in Lean on 28 October 2026; the formal code runs to a reported 1,147 pages.
- The verification process itself is contested: some mathematicians call the closed "shepherding" approach needless gatekeeping, and wider scrutiny of the proof is only beginning.
What Happened on 10 November
On the night of 10 November 2026, a committee of physicists led by Nobel laureate David Gross posted a document explaining how OpenAI's proof works, alongside their affirmation that the proof had been formally verified in Lean, a programming language that checks mathematical work step by step. The committee also released the Lean code OpenAI had submitted, and its own assessment of that code. Scientific American reported the verification on 11 November, under the headline that scientists had confirmed OpenAI's blockbuster quantum physics result.
The announcement matters for two reasons. The first is the result itself: if the committee is right, an AI system has produced a proof of one of the hardest problems in mathematical physics, a problem that has resisted human mathematicians for decades. The second is the way it was verified: by a small committee appointed by OpenAI, working behind closed doors for weeks, using another OpenAI model to help examine the code, and admitting in its own account that it does not yet fully understand the proof. Both parts of the story need to be told, because the second part is what determines how much weight the first part carries.
This article explains what the Yang-Mills problem is, what OpenAI's proof actually establishes, what the committee checked and how, what remains unproven, and why the verification process has divided mathematicians. It keeps the confirmed facts separate from the open questions, because in a story about a proof nobody fully understands, that separation is the whole point.
What the Yang-Mills Problem Is
The Yang-Mills problem is one of the Millennium Prize Problems, a set of seven problems codified by the Clay Mathematics Institute in 2000, each carrying a $1 million prize. Five of the seven remain unsolved. The Yang-Mills problem asks for a rigorous mathematical foundation for quantum Yang-Mills theory — the theory that describes the strong nuclear force, the force that holds atomic nuclei together.
In plain terms, physicists have used Yang-Mills theory successfully for decades. It is a vital foundation for the Standard Model of particle physics, and its predictions match experiments. But using a theory and proving that it exists as a rigorous mathematical object are different things. The Millennium Prize problem asks for the second: a proof that a quantum Yang-Mills theory exists, satisfying a set of axioms known as the Wightman axioms, and a proof that it has a "mass gap" — a minimum energy for its particles, which is what makes the strong force act over short distances rather than infinite ones.
The problem therefore has two halves. The first half is existence: show that the theory can be constructed. The second half is the mass gap: show that the constructed theory has the property that makes it physically meaningful. OpenAI's proof addresses the first half. The second half remains open, and it is the harder half. That is why the prize is not won, and why the headlines saying OpenAI "solved" the Yang-Mills problem need the qualification that follows.
What OpenAI Did, Step by Step
The timeline of OpenAI's work is documented in the committee's account and in the reporting around it, and it is worth laying out in order because the dates matter.
OpenAI says its models produced a written Yang-Mills existence proof on 9 September 2026, and a proof covering quantum chromodynamics — the version of the theory with quarks, the particles that make up protons and neutrons — on 18 September. The company completed the Lean formalisation of the Yang-Mills result on 16 October, and on that day asked David Gross to take custody of the proof and manage its release. OpenAI submitted the proof in Lean on 28 October. The submission is reported to run to 1,147 pages.
In late October, OpenAI announced publicly that its model had solved the first part of the problem — the existence problem. That announcement came ahead of the verification: the claim was public before the committee had finished its work. The company's claim also cut ahead of another mathematician's public, ongoing collaboration with an employee of OpenAI's rival Anthropic to prove the same result, according to Scientific American. The race to the result, and the race to announce it, were both part of the story before the verification was complete.
The announcement in late October was also part of a larger pattern. OpenAI had released more than 700 AI-generated proofs in a single day in October, according to the reporting around the Yang-Mills claim. The Yang-Mills proof was the most significant of those, but it arrived in a context where OpenAI was demonstrating the scale of its models' mathematical output, not only a single result. That context matters for how the claim was received: a company that has just published hundreds of proofs is making a different kind of claim from a mathematician who has spent years on one.
A second OpenAI submission, dated 4 November, is listed as a work in progress. The version with quarks — the quantum chromodynamics proof from 18 September — has not yet been assessed. The verification published on 10 November covers only the first submission, the pure Yang-Mills existence proof of 28 October.
Who Verified It, and How
OpenAI did not publish the proof itself. Instead, the company enlisted the theoretical physicist David Gross to lead a committee to supervise the release. Gross is a Nobel laureate, and he selected the other four members of the committee: Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. Jaffe is a co-author, with Edward Witten, of the official Clay Mathematics Institute description of the Yang-Mills problem, which gives the committee a direct connection to the problem as it was originally posed.
On 9 November, the committee issued a statement previewing its findings, saying that its work confirms that OpenAI's result "proves the existence of pure Yang-Mills theory according to the problem statements which are standard in physics and mathematics". On the night of 10 November, Gross and his four colleagues posted a document explaining how the proof works, alongside their affirmation of the proof's formal verification in Lean. The group said it would release two GitHub repositories on Tuesday 10 November: one containing the full Lean code submitted by OpenAI on 28 October, and the other containing a validation report concluding that the submission proves the existence of pure Yang-Mills theory under standard physics and mathematics problem statements.
The verification had two layers, and the distinction between them is the key to understanding what was actually confirmed. The first layer is formal verification in Lean. Lean is a programming language and proof checker: a proof written in Lean is checked by the computer, step by step, and either every step follows from the previous ones or it does not. If Lean accepts a proof, the logic is correct — within the formal statement that was written. That is a powerful check, and it is the reason a 1,147-page proof can be verified at all: no human could check that many steps by hand.
The second layer is the one Lean cannot do. Lean cannot tell whether the formal statement captures the physics — whether the thing that was proven in Lean is actually the Yang-Mills existence problem as physicists and mathematicians understand it. That is what the committee checked. Its assessment says the code proves that 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. In plainer language: the committee judged that the formal statement matches the real problem, and that the proof, as formalised, establishes the existence part of it.
The committee's checks, as reported, included a clean rebuild of the code, an axiom audit, and a check described as a Comparator check, all of which passed. The assessment itself runs to 35 pages, prepared by Douglas, Jaffe, Ron Nissim and Fred Rajasekaran, and approved by the Yang-Mills group. That is the document that carries the committee's reasoning, and it is the document that wider scrutiny will now examine.
The Caveat in the Committee's Own Words
The most quoted sentence in the coverage is also the most important. The committee wrote: "We used OpenAI's Astra [model] to assist with the digestion of the supplied Lean code. We do not claim to fully understand the proof yet."
That sentence deserves careful reading, because it is doing two things at once. It is being honest: a 1,147-page formal proof, produced by an AI system, is not something five people can fully understand in a few weeks, and the committee says so plainly. It is also a limit on what the verification means. The committee has verified that the formal proof checks out in Lean and that, in its judgement, the formal statement captures the problem. It has not claimed to understand every step of the reasoning in the way a mathematician traditionally understands a proof. Scientific American describes the committee's explanatory document as too cursory to convey how the proof works.
That is not a reason to dismiss the verification. Formal verification in Lean is a real check, and the committee's judgement about the statement is a real judgement from people qualified to make it. But it is a reason to treat the verification as the beginning of scrutiny rather than the end of it. The proof is now public, in Lean code, and the wider mathematical community can examine it. Whether the committee's judgement holds up under that scrutiny is a question that will take months, not days, to answer.
What Was Not Proven
The limits of the result are stated clearly in the committee's own assessment, and they are as important as the result itself.
The proof does not establish a mass gap. The mass gap is the second half of the Millennium Prize problem, and it remains unresolved. Because the prize requires both the existence proof and the mass gap, the $1 million prize is not in play. The working group that reviewed the proof has said explicitly that its finding is not a ruling on the Clay Mathematics Institute's prize, and that it has no current connection with the institute.
The proof also does not prove that the infinite-volume limit is unique. That is a technical point, but it matters: the construction works as a scaling limit, and the question of whether that limit is unique — whether there is only one theory that results — is left open. The remaining questions listed in the reporting include a unique infinite-volume limit, a uniform mass gap, and the Wilson-loop area law. Those are not footnotes. They are the parts of the problem that remain, and they are the parts that would be needed for a complete solution.
The version with quarks has not been assessed. OpenAI says its models produced a quantum chromodynamics proof on 18 September, but the committee's assessment covers only the pure Yang-Mills submission of 28 October. A second submission dated 4 November is listed as a work in progress. Anyone claiming that the full theory of the strong force, with quarks included, has been proven is ahead of the verification.
And the full written proof has not been released. The Lean code and the committee's assessment are public, but the full 1,147-page written proof, in the form a human mathematician would read, has not been published. The committee said it would publish the proofs with its assessment when it was satisfied with their quality and correctness. That is a further reason the wider community's examination is only beginning.
Why the Process Has Divided Mathematicians
The verification is not only a scientific event. It is also a dispute about how science should be done, and that dispute is part of the story.
The committee's "shepherding" approach — a small group appointed by OpenAI, working behind closed doors, releasing the result on its own timetable — contradicts recommendations for transparency from an OpenAI-affiliated advisory group, according to Scientific American, and it has riled some mathematicians who have seen it as an act of needless gatekeeping. The complaint is straightforward: a proof of this importance should be published openly and examined by the whole community, not held by a committee that decides when and how it is released. The committee's reply, implied in its approach, is equally straightforward: a 1,147-page formal proof needs careful handling, and a premature release of an unverified claim would be worse than a delayed release of a verified one.
Both positions have merit, and the dispute is not going to be settled by this article. What can be said is that the process was unusual. OpenAI approached Gross, Gross picked the other members, and the committee's judgement now faces wider scrutiny. That is a different process from the ordinary one, in which a proof is submitted to a journal, refereed anonymously, and published with the reasoning available for anyone to check. The Lean code is now public, which is a substantial step toward openness, but the process that got it there was controlled by the company that produced the result and by the committee it appointed.
There is also a question about the use of OpenAI's Astra model to help examine the code. The committee is open about it, and using an AI model to digest a large formal proof is a reasonable thing to do. But it adds a layer of circularity that the wider community will want to examine: an OpenAI model helped to verify a proof produced by an OpenAI model, under the supervision of a committee appointed by OpenAI. That does not make the verification wrong. It does mean the verification is not independent in the way a fully external check would be, and the community's scrutiny will need to take that into account.
What This Means for AI and Mathematics
If the committee's judgement holds up under wider scrutiny, this is among the hardest results an AI system has produced. That is the assessment in the reporting, and it is a fair one. The Yang-Mills existence problem is not a puzzle or a benchmark. It is a problem that has resisted the best human mathematicians for a quarter of a century, and it sits at the foundation of the Standard Model. An AI system producing a proof of it — even a proof that humans do not yet fully understand — is a genuine milestone in what AI can do.
It is also a milestone that raises a question the mathematics community has been circling for years: what does it mean to prove something if no human understands the proof? A formal proof checked by Lean is correct in a precise sense. But mathematics has traditionally been a human activity, in which a proof is not only correct but understood, explained, and built upon. A proof that is verified but not understood is a new kind of object, and the community has not yet decided what to do with it. The committee's own admission — "we do not claim to fully understand the proof yet" — is the honest starting point for that discussion.
The result also matters for the competition between AI labs. OpenAI's claim cut ahead of a public, ongoing collaboration between another mathematician and an employee of Anthropic to prove the same result. The race to the Yang-Mills proof is, in that sense, part of the wider race between the frontier labs — the same race visible in OpenAI's DevDay 2026 announcements — and the verification is part of that race too. Our earlier analysis of OpenAI and Anthropic's trillion-dollar compute bill explains the scale of the investment behind that race, and the Yang-Mills result is one of the things that investment is now producing.
For readers who are not mathematicians, the practical takeaway is simpler. This is a real result, verified by a serious committee, with clear limits. It proves the existence half of a famous problem. It does not prove the mass gap. It does not win the $1 million prize. It has not yet been fully understood, even by the people who verified it. And the wider community has only just begun to examine it. Those five sentences are the honest summary, and any account that drops one of them is misleading.
What We Could Not Verify
- The full 1,147-page written proof has not been published, so we have not read the proof itself. Our account relies on the committee's assessment and on reporting from Scientific American, Dataconomy and others.
- The committee's explanatory document has been described by Scientific American as too cursory to convey how the proof works. We have not independently assessed that document against the Lean code.
- Whether the committee's judgement will hold up under wider scrutiny cannot be verified yet. The proof has only just been released, and the mathematical community's examination is beginning.
- The claim that the proof runs to 1,147 pages is reported, not independently verified by us against the released code.
- We could not verify the contents of the second OpenAI submission dated 4 November, which is listed as a work in progress, or the quantum chromodynamics proof of 18 September, which has not been assessed.
Sources
- Scientific American, "Scientists confirm OpenAI's blockbuster quantum physics result," 11 November 2026: https://www.scientificamerican.com/article/openais-shepherds-affirm-blockbuster-physics-result/
- Dataconomy, "Nobel Panel Validates OpenAI's Yang-Mills Existence Proof," 12 November 2026: https://dataconomy.com/2026/11/12/nobel-panel-validates-openais-yang-mills-existence-proof/
- Integuide AI News, "OpenAI on fading CoT monitorability, Yang–Mills proof validated," 12 November 2026: https://news.integuide.com/archive/d8e3f23b-39b7-48ac-ac85-97d82567e991
- Pivot News, "OpenAI-appointed committee says AI proof of Yang-Mills existence…," 13 November 2026: https://pivotnews.ai/five/openai-appointed-committee-says-ai-proof-of-yang-mills
- Yang-Mills Group, "Validation of the OpenAI Yang–Mills construction," assessment of the 28 October 2026 submission, 10 November 2026: https://github.com/yang-mills-group/YM