A formal claim with an important caveat
Software developer Dan Abramov says he has produced a Lean proof of a conjecture posed by mathematician John Conway about omnific integers, after spending roughly a month working with an AI model. The result was published on September 18, but it should not yet be treated as a settled mathematical breakthrough: Abramov explicitly says independent mathematicians have not verified it.
The submitted proof has passed mechanical checks in the Palomar registry, according to Abramov's account. People familiar with Lean and the relevant field also told him that the formal statement appeared plausible. Those checks are meaningful because Lean's kernel verifies that each proof step follows from declared assumptions, but they do not by themselves establish that the encoded statement perfectly captures Conway's intended conjecture or that the development avoids every problematic assumption.
What the conjecture concerns
Conway's refinement conjecture addresses factorisation among omnific integers, a class within the surreal-number system. In simplified terms, it proposes that when two products are equal, their factors can be refined into four components and recombined across the equality. Ordinary integers display an analogous property: two different factorisations of the same number can be broken into smaller factors that account for both sides.
Surreal numbers extend far beyond familiar integers and real numbers, encompassing infinitely large and infinitely small values. Their construction begins by placing new values between sets of previously created numbers and continues through transfinite stages. Omnific integers form an integer-like part of that much larger system, so properties that appear routine for finite integers can require substantial proof in this setting.
Abramov selected the problem after asking Claude to suggest an open question in surreal-number research. He then used the model during the effort to formalise and prove the claim. His account also records an early mistake by the model about how prior mathematical work reduced the problem, underscoring that the system's suggestions still required checking and correction.
Why verification remains essential
Formalisation can make a proof easier to audit because a proof assistant checks its internal logical steps. It does not eliminate the need for expert review. Mathematicians still need to examine whether definitions align with the established literature, whether imported results and assumptions are appropriate, and whether the claimed theorem matches the historical conjecture.
The episode is therefore best understood as a reported result and an invitation for scrutiny. If specialists confirm both the formulation and proof, it could demonstrate how AI-assisted exploration and proof assistants can help non-specialists engage with difficult research problems. Until that review occurs, the strongest supported conclusion is narrower: a mechanically accepted Lean artifact exists, and its author believes it proves the conjecture while openly acknowledging that the claim remains unverified.
