At 6 P.M. EDT on October 6, OpenAI dropped 719 mathematical manuscripts spanning 372 problem families onto GitHub, all produced by an internal model. Just one day later, the repository got its first real-world stress test — not a new batch of results, but the first wave of retractions and repairs. The repo's history page records it plainly: one sign error, three papers withdrawn.
Which Three Papers Were Pulled
According to the repository's history page, the error sat in "Algebraicity of Weil classes on split abelian eightfolds": a sign error invalidated a stabilization-trace cancellation argument, and the construction it supported was reused by two dependent papers. So "Algebraicity of Kuga–Satake Correspondences for K3 Surfaces" and "The rational Hodge conjecture for products of K3 surfaces" were withdrawn alongside it. The retracted papers were not quietly deleted — each keeps a README explaining the gap and linking to the archived manuscript, and previously released versions remain accessible.
More Repairs Than Retractions
The same update revised 14 manuscripts: Lipschitz heights and Ashkin–Teller currents (4 papers) repaired crossing, boundary-attachment and convergence arguments; Kähler minimal model programs and abundance (6) expanded positivity and contraction arguments; taming and hypersymplectic deformation (2) corrected a cone-equality claim and added a strict-inclusion example; Incompressible Box Transport revised torus-projection and common-clock estimates; and one Birch–Swinnerton-Dyer paper removed an obsolete citation to a deleted supporting manuscript. Thirteen more manuscripts were updated to cite the revised editions. On the formalization side, 6 new Lean proofs and 5 supporting additions pushed top-line formalized results to 300 of 719, about 42%.
Why This Matters More Than Not Making Mistakes
Scientific American's coverage quoted MIT mathematician Andrew Sutherland: until the model is released and results replicated, claims of one-shotting problems with a single agent should be treated as unverified — "we should ask for receipts." The repo's first maintenance cycle is exactly such a receipt. Its README states plainly that some unformalized results could have issues and that OpenAI will endeavor to fix them quickly — and a day later it did, with version notes attached. Per the README, each result consumed on average three hours of ChatGPT Pro-level thinking compute, drawn from roughly 4,000 posed problems; meanwhile more than half of top-line results still lack Lean formalization as a safety net, so further corrections are nearly inevitable — this wave probably won't be the last.
So What
In traditional mathematics, the road from "an error is found" to "a paper is formally retracted" is often measured in years. This repository offered a different answer within a day: the surprising thing is not that errors exist, but whether there is infrastructure to make them traceable, archived and version-controlled. The Apache-2.0 license, per-paper citation formats, and version-preservation policy turn a retraction from a scandal into an ordinary commit. The credibility of AI-generated knowledge will ultimately depend not on models never erring, but on whether correction speed and transparency can keep up with generation speed. That is a lesson every AI team dumping machine output onto GitHub should copy.