OpenAI brought three giant proofs to the lab door. Show the verification ledger

THE SIGNAL IN ONE SENTENCE

Three enormous scientific claims arrived together on Monday, which is one way to make a quiet week impossible. OpenAI said its internal models produced new work in harmonic analysis, computational complexity and quantum field theory. The company says one result proves the local smoothing conjecture in every dimension, with consequences for the Kakeya and Stein restriction conjectures. A second says the Exponential Time Hypothesis is false. A third claims mathematically rigorous existence results for quantum field theories that include quantum chromodynamics and a completion of the Standard Model. Those are not tidy product updates. They touch problems that have organized years of human research. They are also claims from the organization that built the models, not three settled conclusions delivered by the scientific community. OpenAI makes that boundary unusually visible in its October 26 announcement. It calls the manuscripts a form of raw research data and says the release process is an experiment. That framing is the most important part of the news. The plain signal is this: AI can place a candidate proof on the table. Science still needs a public ledger showing what was claimed, what was formalized, what experts understand, what remains hidden and what the wider field has independently accepted. OpenAI is using a different disclosure path for each result. For the local smoothing result, the company released a natural-language manuscript and a Lean formalization. OpenAI says the model produced the output on October 18 and that it was formalized the next day. Its public Lean scope page says the development covers local smoothing estimates in every dimension, related multiplier and extension estimates, a Kakeya maximal estimate and full Hausdorff dimension for Kakeya sets. That is a serious artifact. It is not a magic stamp labeled true. Lean is a proof assistant. It checks whether a formal proof follows from the definitions, statements and trusted foundations encoded for it. That can catch gaps that ordinary prose hides. It can also verify exactly the theorem that was entered while leaving humans to check whether the formal statement matches the intended informal claim, whether the assumptions capture the scientific problem and whether the argument contributes the understanding researchers need. The Exponential Time Hypothesis claim has a second path. ETH is a widely used conjecture about the time required to solve Boolean satisfiability problems. OpenAI says its system found a natural-language disproof on October 16 and that a Lean formalization followed on October 17. The public formalization scope describes an exact counting and satisfiability algorithm with subexponential time for fixed clause width under the encoded conditions. OpenAI has released the natural-language writeup and Lean materials. It has also asked a group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams to understand and explain the result. Impagliazzo and Paturi helped introduce ETH. Williams has spent years studying algorithms and lower bounds around it. Their participation is relevant evidence about the seriousness of the review effort. It is not an endorsement announced in advance. OpenAI says the group will verify and shepherd the work, that all raw result data will be public, and that the company will fund a conference on the implications. The announcement does not contain a final report from that group. Until the experts publish one, the careful sentence is that OpenAI has released an AI-generated disproof claim with a formal artifact and an independent examination underway. The physics result is on a third path, and it has the widest gap between claim and public evidence. OpenAI says its models produced a natural-language proof for Yang-Mills existence on September 9 and one for quantum chromodynamics on September 18. It says the Yang-Mills formalization finished on October 16. A group led by Nobel laureate David Gross, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov, will examine and explain the work. Unlike the mathematics and complexity results, those materials are not public yet. OpenAI says it deferred to the group, which chose to write up the work before release. That may be a reasonable way to prevent a large, technical artifact from being mistaken for a finished scientific account. It also creates an information asymmetry. A small invited group can inspect evidence that the rest of the field cannot. OpenAI acknowledges that choosing initial shepherds shapes who gets involved first. The useful response is not to flatten the three claims into one victory lap. Give each one its own verification ledger. Start with the claim itself. Write the exact theorem in ordinary language and formal notation. Record the assumptions, the model-visible context, the date the candidate proof appeared and every later revision. Then record the artifact layer. Is the natural-language proof public? Is the formal statement public? Can outside researchers run the proof checker from a pinned environment? Does the repository show dependencies, compiler versions, trusted axioms and the difference between the manuscript and the formal theorem? Next comes the understanding layer. Name the experts who examined the work, their access, their conflicts and whether they were paid. Publish their conclusions in their own words. A correct formal object can still be difficult to connect to existing literature, teach to students or turn into better methods. Finally, record community uptake. Have independent groups reproduced the formal check? Have specialists found a shorter argument, a missing assumption or a link to prior work? Has the result survived workshops, seminars and ordinary adversarial reading? Have later papers used it successfully? That last layer cannot be rushed by a press release. Science is slow partly because many forms of error look impressive for a while. Credit needs the same care. OpenAI says it will not seek academic credit for discoveries produced by its internal models. It says external groups should receive recognition for the understanding and writeups they create, and it commits to crediting prior researchers and ideas. Those are useful principles. They still leave hard questions. Who is responsible for the choice of problem, the prompt context, the model run, the selection of a promising output and the decision to announce it? Who deserves credit when a human identifies the missing bridge that turns a machine artifact into a theorem other scientists can use? How should a journal list a model, a laboratory and an independent verifier when their contributions are different kinds of work? A verification ledger will not settle authorship. It can keep the contributions from collapsing into one glossy noun called discovery. There are practical applications here even before any theorem changes a textbook. Research groups can use the same layered process for smaller AI-generated claims. A materials lab can separate a proposed mechanism from the simulation that supports it and the experiment that tests it. A software team can separate an automatically generated invariant from the proof assistant check and the behavior of the deployed system. A policy office can separate a model's legal interpretation from the cited authority and the qualified human decision. The pattern is portable because it is humble. Machines are good at producing more candidate work than people can comfortably review. That makes verification capacity a bottleneck. The answer is not to lower the bar. It is to design better queues, clearer artifacts, funded review and public revision histories. OpenAI says its external collaborations are voluntary and unpaid, while the company is also supporting broader independent work through funding. The practical test will be whether reviewers have enough time, tools and freedom to disagree publicly. Independence without resources can become a polite invitation to donate scarce expertise. Watch the next documents, not the next adjective. For the local smoothing claim, look for independent mathematicians to compare the Lean statement with the intended conjecture and explain the new auxiliary partial differential equation. For ETH, look for the shepherding group's report, reproducible formal builds and the community's response to the claimed algorithm. For the physics work, look for public artifacts, an exact statement of assumptions and a clear account of what the construction proves relative to the Clay Millennium problem. Also watch for corrections. A revision is not evidence that the whole exercise failed. A hidden revision is. OpenAI has put extraordinary claims into three different verification pipelines. That is more informative than pretending every claim is already a discovery, and more useful than dismissing every machine-produced proof on sight. The right posture is curiosity with paperwork. Open the artifacts. Pin the theorem. Name the reviewers. Fund the checking. Publish disagreement. Keep the revision trail. Let the word breakthrough arrive last.