Lean FRO and Mathlib Certify Riemann Hypothesis Proof: Four Independent Checks Clear in 24 Hours
Alpöge's GRH proof in Lean, built with Claude, cleared nanoda, con-leche, and lean4lean without errors.

One of mathematics' greatest unsolved problems has cleared its first major verification hurdle. On November 20, 2026, Levent Alpöge—a mathematician at Anthropic who developed the proof working with Claude—posted an enormous Lean 4 formalization of the Generalized Riemann Hypothesis (GRH) to a dedicated public GitHub repository. Within hours, four independent proof-checking programs had accepted the code without errors. The following day, the Lean Focused Research Organization (Lean FRO) and the Mathlib community published a joint statement confirming that the proof had cleared every criterion on Kevin Buzzard's established four-point verification checklist for major Lean formalizations.
The Generalized Riemann Hypothesis is a sweeping statement about the distribution of zeros of a broad class of mathematical functions called L-functions, of which the Riemann zeta function is the most famous example. The classical Riemann Hypothesis—the precise statement that Bernard Riemann conjectured in 1859 and that the Clay Mathematics Institute lists among its seven Millennium Prize Problems—follows as a consequence. If the GRH proof withstands sustained peer review, it would resolve not only the 167-year-old Riemann conjecture but a family of related conjectures that underpin large portions of modern number theory.
Terence Tao, the UCLA Fields Medalist, described Alpöge's approach on his blog as relying on "the full power of modern developments on the Langlands program and arithmetic geometry," and he noted that digesting the mathematics will "take some time." Roughly five weeks before the announcement, Tao had told Brady Haran in a Numberphile interview titled "No More Heroes" that real AI-assisted progress on the Riemann Hypothesis might not arrive until "some combination of humans and AIs in 2050 or something"—a timeline the events of November 20 compressed by roughly two and a half decades.
Read more: Anthropic mathematician claims proof of Riemann Hypothesis built with Claude
What the Four-Check Standard Requires
Kevin Buzzard, a professor at Imperial College London and founder of the Xena Project, laid out the benchmark in a November 12, 2026, blog post titled "…and the proof is checked in Lean." A Lean repository earns full confidence only when it clears four distinct questions: Does the Lean statement accurately express the intended informal theorem? Does the code compile? Does the code actually prove the stated theorem, without hidden redefinitions or extra axioms slipping in? And does the proof exploit no bug in Lean's proof-checking kernel?
Alpöge's repository passed all four. Andrew Sutherland, a number theorist and professor at MIT, ran and documented the independent checks on a dedicated verification page that became the primary public record of the certification. Sutherland confirmed that the Lean statement of the GRH matched the standard informal statement, that the code compiled successfully, and that two separate comparison tools—including the paranoid lake comparator --paranoid mode—found no hidden redefinitions or extra axioms. Sutherland also ran an axiom audit: the proof relies only on three standard Lean axioms (propositional extensionality, classical choice, and quotient soundness), with no non-standard assumptions introduced.
Three Kernels, No Errors
The deepest layer of the check was kernel-level: running the proof through software that re-derives its validity from first principles, entirely independent of Lean's standard proof checker. Three such programs each accepted the proof:
Nanoda, a Lean proof checker written in Rust by Austin Theriault and maintained under a permissive license, completed its verification in 27 minutes and 42 seconds running on 80 CPU threads, exiting with code 0.
Con-leche, the Lean FRO's own implementation of the Lean kernel written in Lean itself, accepted the proof in just over three minutes. Con-leche carries an additional layer of assurance: it includes a machine-checked consistency proof—developed with AI assistance under Joachim Breitner's supervision at the Lean FRO—guaranteeing that any remaining unsoundness in the checking process would have to arise from Lean's compiler or runtime, not from the kernel logic itself. Con-ron, an Aeneas-verified Rust port of con-leche, also runs as part of the paranoid kernel suite.
Lean4lean, an external Lean verifier written in Lean by Mario Carneiro at Chalmers University of Technology—and independently maintained at github.com/digama0/lean4lean—passed 175,187 of 175,187 declarations, completing around 21:55 UTC on November 20. Five modules required a raised step limit, a known performance adjustment rather than a correctness concern.
The bundled con-leche and con-ron binaries shipped with Lean v4.36.0 declined to process one reflexive inductive type in the proof—a limitation of those particular builds—but the current con-leche master branch accepted the full proof without issue.
What This Verification Does Not Establish
Buzzard was careful to draw a sharp line between machine verification and mathematical understanding. In a statement quoted across multiple publications, he said the certification confirmed that the code "definitely proves something" and that the statement "is exactly the standard statement"—while stressing that mathematicians have not yet read the proof. The 17-million-line Lean file, generated through Alpöge's collaboration with Claude, is too large for conventional human review and uses mathematical strategy that Tao described as drawing on the most advanced current machinery in the Langlands program.
The Clay Mathematics Institute's rules for the Riemann Hypothesis Millennium Prize are unambiguous: a solution must appear in a refereed journal of worldwide repute and must achieve general acceptance in the mathematics community for at least two years following publication. Machine certification, however rigorous, does not satisfy either condition. The prize is not triggered.
The Proof's Mathematical Approach
Rather than attacking the Riemann zeta function directly in the complex plane, Alpöge appears to have pursued the problem through a 1984 reformulation proved by the French mathematician Guy Robin. Robin's theorem established that the Riemann Hypothesis is true if and only if the inequality σ(n) < e^γ · n · log log n holds for every integer n greater than 5,040—where σ(n) is the sum of all divisors of n and γ is the Euler–Mascheroni constant. This elementary number-theoretic criterion has long been used as a computational and theoretical handle on the hypothesis. Tao noted on his blog that the full strategy appears to draw on "the full power of modern developments on the Langlands program and arithmetic geometry," suggesting Robin's criterion is one entry point into a much deeper argument.
Community Scrutiny Is Just Beginning
The Lean Zulip server, the primary real-time forum for the Lean proof-assistant community, has been hosting intensive discussion since the announcement. Moderators have asked participants to focus discussion on Lean-technical aspects of the proof rather than mathematical interpretation. Some researchers have raised questions about whether certain parts of the Lean formalization accurately encode the mathematical objects they claim to represent—a standard and expected part of the scrutiny process.
Sutherland's verification page, which continues to be updated as additional checks run, described his own reaction as finding the statistical likelihood that something fundamental is wrong to be "approximately zero"—a phrase that struck a deliberately cautious note, acknowledging the unprecedented scale of AI-generated formal mathematics while affirming the breadth of the verification already completed.
The coming months will determine whether the mathematical ideas inside the Lean formalization can be extracted, understood, and independently confirmed by the broader research community—the step that separates machine-certified from humanly understood.



