The shepherds are still the reviewers
OpenAI says its models produced proofs and asked outsiders to check them. That reviewing role is not a certification. Physics files are held back.
The short version
No. OpenAI asked outside groups to review. The computer-science letter, as reported, says involvement is not certification. The physics group is withholding the model output while it writes its own account.
The post
OpenAI's note on working with academic communities says internal models produced three results: a proof of the local smoothing conjecture in every dimension, a disproof of the Exponential Time Hypothesis, and existence arguments for certain interacting quantum field theories in four dimensions, including Yang-Mills and quantum chromodynamics. The company calls the manuscripts raw research data and says it asked outside experts to shepherd each one. The Decoder and The Verge reported the arrangement on October 27. A shepherd, in this post, is a reviewer with a plan for how the community sees the work. A shepherd is not a coauthor who has signed the theorem, and OpenAI does not claim that signature.
Lean checks a statement, as written
OpenAI says the mathematics and computer-science arguments were formalized in Lean, and that the computer-science writeup and Lean file are in a public repository. Lean checks that a proof establishes the statement that was typed into Lean. The Decoder's account of the post makes the obvious next point: experts still have to decide whether that statement is the conjecture people meant. Weights a reader can fetch are a different brief. A Lean file a company posted is closer to that idea than a private model card is, and it is still not a certificate from the conjecture's authors.Weights you can fetch.
The computer-science letter
OpenAI dates the natural-language disproof of the Exponential Time Hypothesis to October 16, 2026, and the Lean formalization to October 17. The reviewers it names are Russell Impagliazzo and Ramamohan Paturi, who are associated with the hypothesis, and Ryan Williams of MIT. The Decoder reports their letter: involvement should not be taken as certification, they have not signed a nondisclosure agreement, and they are not being paid. The Verge reports the same group. TokenGate was about when a lab publishes. Here the lab published the claim and, for this result, the files, and the people most entitled to object are on the record that they have not yet objected or agreed.TokenGate's papers.
The physics hold
OpenAI says the Yang-Mills existence argument was produced on September 9, the QCD argument on September 18, and the Lean formalization of Yang-Mills existence on October 16. The shepherd group it names is led by David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov, and Nikita Nekrasov. Unlike the computer-science group, OpenAI says this group chose not to release the model-produced materials yet and will write the account itself. A result that the reviewers are holding is not a result a reader can check. The local-smoothing claim is the third track: OpenAI says the argument was obtained on October 18 and formalized on October 19, and that a proof in every dimension would also resolve the Kakeya and Stein restriction conjectures in all dimensions. That implication is OpenAI's description of the mathematics. It is not a referee report.Nanbeige's table.
The prize, left unclaimed
Yang-Mills existence is one piece of a Clay Mathematics Institute problem that also asks for a mass gap. A later account of the post says the existence claim does not answer the mass gap, so it does not finish the prize problem. This page did not audit the Lean file against the Clay statement. It will not call the blog a prize citation. Terra's addendum, filed beside this, is a product grade and a price. It is not a theorem. The two OpenAI documents this week should stay two documents.Terra's warning rate.
What the company says about the process
OpenAI says the shepherd approach has a cost: the first reviewers learn sooner than everyone else. It says it agreed, for the computer-science claim, to publish the raw data and to fund a conference. It also says it has released large batches of model-written mathematics before, and that proofs without a formal check can contain errors. That last admission is why the Lean step is in the post, and why the Lean step is still not the end. d3-ultra's card withholds its test questions. OpenAI, for two of these three claims, says the files are public. Public files are the better of those two habits. They are not a finished review.The d3-ultra card.
Who is in scope
The people in scope are the mathematicians, computer scientists, and physicists who can read the files, and anyone tempted to treat a model vendor's blog as a solved conjecture. The shepherds OpenAI names are in scope as reviewers. Their universities are not coauthors of the claim. A reader who does not work in these fields has no action item other than to wait for the groups' own writeups.
What this page will not do
This page will not sketch a proof, quote a Lean tactic, or decide whether the Exponential Time Hypothesis is false. The Qwen brief was a measured gap on a public test. These manuscripts are the opposite shape until the shepherds speak: a claim, a formalization the company vouches for, and a letter that says the obvious experts have not certified it. On-device inference is about a phone. It has nothing to do with a Millennium problem, except that both can be overclaimed in a headline.What on-device inference keeps.
The record
The record is OpenAI's post for the claims and the dates, The Decoder for the letter that withholds certification, and The Verge for a second account of the same groups. The physics materials are not public, the company says. The computer-science files are public, the company says. The shepherds are still the reviewers.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
What does OpenAI say its models produced?
A proof of the local smoothing conjecture in every dimension, a disproof of the Exponential Time Hypothesis, and existence arguments for some four-dimensional quantum field theories, including Yang-Mills and QCD.
Which dates does OpenAI give for the computer-science claim?
A natural-language disproof on October 16, 2026, and a Lean formalization on October 17. OpenAI says it posted the writeup and the Lean file.
Who is reviewing the computer-science claim?
Russell Impagliazzo and Ramamohan Paturi of UC San Diego, and Ryan Williams of MIT. Their letter, as The Decoder reports it, says involvement is not a certification.
What did the physics group do with the files?
OpenAI says a group led by David Gross chose to hold the model-produced materials and write the work up itself. Lean formalization of the Yang-Mills existence argument is dated October 16.
Does an existence argument collect the Clay prize?
OpenAI describes existence proofs. A later account says the mass gap, which the Clay problem also asks for, is a separate question. This page will not call the post a prize citation.