A public elliptic-curve leaderboard maintained by an NSF mathematics institute recorded two new rank records within four days, and the commentary on both credits Claude alongside two human mathematicians. Curve #302 carries the note "BSD + GRH certified to rank 31, found by Claude, Levent Alpoge, and Ava Howell," submitted on August 23, 2026. A rank-30 record on curve #273 landed three days earlier, with the same trio credited.
Key facts
- Curve #273 was submitted on August 20, 2026 with rank at least 30; curve #302 followed on August 23 with rank at least 31.
- Both records are credited to Claude working with mathematicians Levent Alpoge and Ava Howell.
- Each submission publishes its witness points -- curve #273 lists 30 independent rational points, some with numerators hundreds of digits long.
- Primary source: the Elliptic Curve Rank Leaderboard run by the NSF Institute for Computer-Aided Reasoning in Mathematics, under grant DMS 2425401.
An elliptic curve is an equation of a particular shape, and its rank counts how many independent rational solutions it has that can be combined to generate infinitely many more. Rank is deeply studied and stubbornly hard to push upward: constructing a curve with a high rank means finding a specific equation whose coefficients run to sixty-odd digits and then exhibiting thirty-plus independent points on it. Whether ranks can grow without bound is itself an open question, which is why the leaderboard's top submitter account is named ranksunbounded.
What makes these entries interesting is not that a computer searched -- computer search has been standard in this area for decades. It is the division of labour recorded in public. The submissions come with explicit witness points, so the claim is checkable by anyone with the right software, and the commentary on curve #273 reads like a working seminar: a note that the original submission silently dropped one of the witness points because of a parser bug, a link to the exact commit that fixed it, an argument that under the relevant conjectures the rank is exactly 30 rather than merely at least 30, and then a human editing another human's comment to add that it was Claude, with Alpoge and Howell.
That is the useful analogy for where AI-assisted mathematics currently sits. This is not a machine handing down a theorem. It is closer to a very fast graduate student running search strategies while two mathematicians decide what to search for and check what comes back -- and the checking is real, because a rank claim is falsifiable by anyone who plugs the published points back into the curve.
Why it matters: mathematics is one of the few fields where an AI contribution can be audited to the last digit, which makes it the cleanest available testbed for claims about machine discovery. It also arrives in a busy week: the Station multi-agent environment published new results on five of twelve open construction problems the same week, following earlier episodes like AlphaEvolve tightening the matrix-multiplication exponent and OpenAI's ten Lean-checked math claims.
The caveats deserve to be stated as plainly as the records. The leaderboard reports rank as a lower bound -- what the witness points prove -- and the exact-rank statements are certified only under the Birch-Swinnerton-Dyer conjecture and the generalised Riemann hypothesis, neither of which is proven. A conditional certification is a genuine mathematical statement, not a hedge, but it is not the same as a proof from nothing. And the parser bug on the first submission is a reminder of the ordinary failure mode here: the mathematics was right, and the pipeline around it quietly dropped a point. As Terence Tao has argued, the bottleneck is understanding rather than proofs -- and a record on a leaderboard is a data point in that argument, not a settlement of it.
It is worth being specific about what a witness point looks like, because the scale is where the difficulty lives. Curve #273's published witnesses include coordinates like a numerator running to more than thirty digits over a denominator of 9, and others with denominators in the hundreds of millions. These are not numbers you stumble onto. Finding thirty of them that are genuinely independent -- none reachable by combining the others -- is the entire game, and it is why high-rank construction has been a computational sport for decades rather than a pen-and-paper exercise.
The publication model around these records is arguably as interesting as the records. There is no press release and no paper. There is a leaderboard entry with the full equation, the witness points, a naive height, a regulator, a discriminant, a submission timestamp, an edit history, and a comment thread where the humans argue and correct each other in public. Anyone can pull the JSON and check the claim in an afternoon. For a field currently drowning in unverifiable assertions about what AI systems have discovered, that is a fairly good template: publish the object, publish the certificate, let the record stand or fall on arithmetic.
Originally published on Ground Truth, where every claim is checked against the primary source.
