A Lean proof certifies the Riemann hypothesis

The Lean and Mathlib community claims to have certified a formalization of the Riemann hypothesis on November 21, 2026. The proof has been verified by several independent kernels and tools, but its mathematical explanation remains to be clarified.

The Lean and Mathlib community certified on November 21, 2026, a formalization of the Riemann hypothesis attributed to Levent Alpöge, after a day of compilation, statement verification, and checks by the proof kernel. The work, presented as a Lean library of 17 million lines, was also associated with Claude in the narrative published by Alpöge the day before. This certification matters because it validates a very large formal artifact, without resolving the fundamental question of human understanding of the proof.

What the community actually certified

The validation concerns a specific object: a formal development in Lean, not a classic mathematical article ready to be read in one go. According to the joint statement of the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative, the system verifies that the statement and the proof indeed correspond to the Riemann hypothesis as defined in Mathlib. In other words, the machine does not “guess” anything; it checks syntax, definitions, and a logical kernel.

In this case, the nuance is central. The certification establishes that a set of modules compiles without error and that independent verifications accept the result, but it does not by itself transform the text into an elegant mathematical explanation. This is precisely the point Levent Alpöge says he still wants to simplify and understand.

Why multiple verifiers change the game

The robustness of the matter comes from the number of independent checks. Andrew Sutherland, in his MIT verification journal updated on November 21, 2026, at 12:55 UTC, reports that 175,187 modules compiled without error with Lean and Mathlib 4.33.1, his verifications having been launched with Claude Code. An audit of the axioms noted only propext, Classical.choice, and Quot.sound, without any axiom added outside the Lean kernel.

Several external verifiers then accepted the result: Nanoda, an independent Rust kernel, as well as con-leche, which also validated the statement comparison. Versions of lean4lean accepted each compiled module. This type of cross-check limits the risk of a simple tool error, as the examined export is inert, thus cut off from source code that could be manipulated.

Limits reported by the verifications themselves

The certification does not close all doors. Sutherland specifies that some Lean kernel verifications were still ongoing at the time of his report, as well as the complete lake comparator, and that old binaries of con-leche and con-ron exited without verdict due to an unsupported reflexive inductive type. He also indicates that the build ran on non-sandboxed cloud machines and that the export hash was not compared to a sandboxed compilation. No panic: the proof is not “invalid,” but its trust chain still has some bolts to tighten.

Another methodological point: a “passed” result depends on the kernels used and on whether Mathlib accurately represents the targeted mathematics. The system thus verifies a rigorous formalization, not a philosophical truth fallen from the sky.

A Riemann hypothesis in Mathlib format

The verified theorem follows Mathlib’s formal definition: every zero of the Riemann zeta function, except trivial zeros at negative even integers, must have a real part equal to one half. The condition s ≠ 1 appears because Mathlib assigns the zeta function a conventional value at its pole. The repository also verifies statements of the generalized Riemann hypothesis for Dirichlet L-functions on all moduli and characters, but explicitly excludes Dedekind zeta functions, Hecke L-functions, and general automorphic L-functions.

Claude, from assistance to discovery

In Alpöge’s narrative, Claude is not limited to formalization. The mathematician explains that the system accompanied him from the “toy setting” to a more ambitious analogous case, before surpassing him near the end. He even cites a technical example: Claude reportedly proposed that a von Mangoldt sum of length X use a chain of theta lifts of length X. This kind of detail gives an idea of where software assistance can weigh in a long proof, not just in its formatting.

The essential point remains: the certification concerns a formal artifact, not yet an accessible mathematical exposition. Alpöge says he is working to simplify the argument and explain it properly. At this stage, the community thus knows that a Lean proof stands; it does not yet know if it will also hold on a page readable by hurried humans.

What this matter changes for mathematical research

The episode shows that AI-assisted proofs can go beyond the simple status of an intelligent draft. Here, they enter the zone of verifiable artifacts, inspected by several kernels and independent tools. For research, this changes the relationship between assistance, formalization, and trust: one no longer only asks a model to propose an idea, but to produce an object that other verifiers can audit. And that, in formal mathematics, is not a minor detail.

The real test begins now: to understand, condense, and disseminate the argument without losing rigor. It is often there that great proofs cease to be compilation victories and become results that mathematicians truly appropriate.

Key points

  • November 21, 2026: certification by Lean and Mathlib.
  • 175,187 modules compiled without error, according to Andrew Sutherland.
  • Levent Alpöge published the initial announcement on November 20, 2026.
  • Claude participated in the discovery and formalization.
  • Nanoda and con-leche accepted the proof export.
  • Mathlib excludes Dedekind, Hecke, and automorphics from the verified scope.

By the numbers

  • 17 million lines — announced size of Alpöge’s Lean library.
  • 175,187 modules — successful compilation worldwide, November 21, 2026, according to Andrew Sutherland.
  • 5 theorems — dependency noted by the axiom audit.
  • 4 verifiers — Lean, con-leche, con-ron, and Nanoda cited in the checks.

To read