New checks test Claude-assisted Riemann hypothesis proof in two equivalent forms

A checker accepted all 175,187 compiled modules in the Claude-assisted proof claim. Separate tests checked two equivalent arithmetic formulations.

Black-and-white portrait of Bernhard Riemann wearing glasses and a dark jacket
Archival portrait of German mathematician Bernhard Riemann, dated 1863. Smithsonian Institution Libraries. No known copyright restrictions. Cropped and resized.

An additional software checker has accepted all 175,187 compiled modules in Levent Alpöge’s Claude-assisted Riemann hypothesis claim. Separate tests also covered two equivalent ways to state the problem in terms of whole numbers. In verification notes updated at 10:05 a.m. UTC on Nov. 21, MIT mathematician Andrew Sutherland reported that the released Lean source compiled and that two independently written checkers accepted the exported proof. Further checks of the main export were still in progress at that timestamp.

The results make the released files materially easier to assess. They include evidence that the code builds, that separate programs accepted its formal deductions, and that a comparison check matched the statements to the challenge. Those answers are distinct from the work of producing a readable mathematical explanation or meeting the Clay Mathematics Institute’s prize rules.

What the independent Lean checks found

QuestionStatus in Sutherland’s Nov. 21 report
Can the released source compile?All 175,187 modules compiled and passed module-level Lean kernel replay.
Can separate checkers accept the exported proof?Nanoda and con-leche accepted the full export.
Do the statements match the challenge?A separate comparison run matched all five target statements. The end-to-end Comparator run and further Lean replay of the main export were still running.
Status as reported by Andrew Sutherland at 10:05 a.m. UTC on Nov. 21. Some checks were still underway.

Compilation lets readers know the source can be processed in the specified Lean environment. Kernel replay asks the program to validate the deductions in the compiled modules. A separate checker receives an export and examines formal proof data using its own implementation. Using more than one checker gives researchers more evidence about whether the result depends on a single implementation’s behavior.

The statement comparison matters because a computer checks the proposition it is given. A sound derivation would be beside the point if its conclusion were a different proposition. In this case, Sutherland identified the main theorem as Mathlib’s formulation of the Riemann hypothesis and reported that the independently implemented comparison accepted all five challenges. The broader claims in his report concern Dirichlet L-functions.

Sutherland says he ran the checks with Claude Code and published logs and files for others to inspect. The Clay Mathematics Institute’s public list still shows the Riemann hypothesis among the unsolved Millennium Prize Problems. Under its rules, consideration comes after publication in a qualifying outlet, at least two years after publication, and general acceptance among mathematicians. These are institutional milestones alongside the mathematical and software checks.

Mathematician Levent Alpöge said Friday that he believes he and Anthropic’s Claude AI found a proof of the Riemann hypothesis, an 1859 conjecture that describes a deep relationship between the zeros of a mathematical function and the distribution of prime numbers. In a note on his personal website, Alpöge said he posted a formalization in the Lean proof language and plans to write an explanation that mathematicians can read.

What the additional checks found

In a Nov. 21 update posted at 10:05 a.m. UTC, MIT mathematician Andrew Sutherland reported two more ways he checked the work. A checker called lean4lean accepted all 175,187 compiled modules. He also reported successful checks on two equivalent formulations of the Riemann hypothesis phrased in terms of the divisors of whole numbers. The results add evidence about the formal argument and about which mathematical statement it proves.

The module count includes a detail that helps explain how to read machine-verification reports. Sutherland says 175,182 modules passed with the checker’s usual settings. Five reached its default step limit and passed when rerun with that limit raised. The log identifies the initial outcome as a timeout. The completed reruns are why the final count includes all five modules.

Why verify an equivalent statement?

One check asks whether a machine can retrace the argument from its stated assumptions. Another asks whether the conclusion encoded in the software corresponds to the question mathematicians know as the Riemann hypothesis. Sutherland’s arithmetic checks address the second question from an additional direction. He combined the proof with a separately written Lean formalization of criteria due to Guy Robin and Jeffrey Lagarias. The resulting statements concern sums of divisors. For example, the positive divisors of 12 are 1, 2, 3, 4, 6 and 12, and their sum is 28.

Robin’s and Lagarias’s criteria give precise bounds on such sums. Formalizing the equivalences gives reviewers a route from the claimed result about the zeta function to consequences stated in arithmetic language. There are still shared dependencies, including the definitions in the mathematics library and the formal proofs of the equivalences. Sutherland reports that multiple checkers accepted the combined arithmetic proof, including the run using the repository’s own verification configuration. He also distinguishes those results from a separate run on the much larger combined export that was still underway.

The submission note and the verification record let readers track the artifacts separately. One is the compiled set of individual modules. Another is a full export of the main proof. A third combines the argument with the arithmetic equivalences. Sutherland reported passes on particular checks of each artifact. At the 10:05 UTC snapshot, the end-to-end Comparator run and Lean-kernel replay of the main export were still in progress. A pass on the arithmetic check should be read as a result about that combined proof.

What the Riemann hypothesis says

There are infinitely many prime numbers. Their gaps are irregular, but mathematicians have ways to estimate how many primes lie below a given number. The Riemann hypothesis addresses the accuracy of those estimates through an apparently different object called the zeta function. Its zeros are inputs that make the function equal zero. The important ones for this problem are called the nontrivial zeros.

Each nontrivial zero has a real and an imaginary part. The conjecture says the real part is always one-half. On a graph of complex numbers, that means every nontrivial zero lies on the same vertical line. A proof would settle a question about the whole infinite collection and firm up results in number theory that currently depend on the conjecture.

Alpöge also says his approach extends to the generalized Riemann hypothesis, which asks about a broader family of related functions. That claim needs review on its own terms. In the same note he describes a further extension as something he expects with more time. Those are separate statements about the reach of the work.

How Friday’s claim differs from the August result

In research published Aug. 10, Anthropic reported that an unreleased research version of Claude raised a lower bound for the share of zeros on the critical line from 41.6% to 67.2%. The increase was 25.6 percentage points, calculated by subtracting the previous bound from the new one. The company characterized that finding as progress on a related problem.

A lower bound tells mathematicians the minimum share that the argument establishes. The hypothesis itself is about every nontrivial zero. A claim about all of them requires an argument of a different scope, which is what Alpöge now says he and Claude have found. The comparison gives readers a way to distinguish references to August’s documented result from discussion of Friday’s proposed proof. Our earlier reporting on METR’s assessment of Anthropic’s internal benchmarks addresses a separate measure of AI progress.

What happens to the $1 million prize

The institute’s Millennium Prize rules give the prize its own timeline. Before considering a proposed solution, Clay requires publication in a qualifying outlet, at least two years after that publication, and general acceptance within the mathematics community. The clock described in those rules runs from qualifying publication.

The next evidence to watch is a readable argument, the results of the remaining checks, and substantive assessments from specialists in number theory. Those steps would answer different questions about the same claim. Clay’s published status and prize rules give readers a separate public record to follow as the work is evaluated.

Trust & transparency

How we handled this story

Verified

Compared the live article’s earlier snapshot with Andrew Sutherland’s 10:05 UTC record, including the lean4lean totals, arithmetic comparators and still-running checks. Rechecked the Clay Mathematics Institute problem status.

Grounded in this reporting

Ask Newsmerica

Ask about this article. Answers quote the closest published passages and link back to their sources.

This tool retrieves article passages; it does not add outside facts.

Keep exploring

More from Newsmerica

Leave a Reply

Your email address will not be published. Required fields are marked *