---
title: 'In the News: November 21, 2026 (Extra 2)'
description: 'The Lean and Mathlib communities certified a 17-million-line Lean proof of the Riemann Hypothesis that Levent Alpöge says he built with Claude.'
canonical_url: 'https://darkfactory.dev/news/2026-11-21-extra-2'
markdown_url: 'https://darkfactory.dev/news/2026-11-21-extra-2.md'
collection: news
date_published: '2026-11-21T16:01:00-05:00'
date_modified: '2026-11-21T16:01:00-05:00'
---

# In the News: November 21, 2026 (Extra 2)


The Lean and Mathlib communities say a proof of the Riemann Hypothesis is correct. The proof is a Lean library of 17 million lines that the mathematician Levent Alpöge says he wrote with Claude, and it comes with no human-readable paper. The verdict rests on machine checks. For anyone verifying agent output they will never review line by line, the useful part is how those checks were set up.

## 1. Lean community certifies a 17-million-line Riemann Hypothesis proof built with Claude

**[Lean verification of the Riemann Hypothesis](https://leanprover-community.github.io/blog/posts/riemann-hypothesis/)** · Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative · Lean community blog, November 21, 2026

On November 20, Alpöge posted a short [remark](https://alpo.ge/remark.pdf) saying that "working with Claude" he had found a proof of the Riemann Hypothesis and of the Generalized Riemann Hypothesis, and that the two had "formalized the arguments without assumptions." He released no human-readable paper. He writes that the argument is "incredibly long at the moment" and that "I decided to only present Lean." He also says he has not yet mastered all of the mathematics in it and expects the argument to get shorter.

A day later the joint statement says: "We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct." The theorem being proved is Mathlib's existing, expert-vetted definition. The community wrote a separate file that states the theorem and imports only Mathlib, then confirmed with two independent tools that Alpöge's library proves exactly that statement. According to the statement, several checkers have checked the exported raw proof, among them Lean's own kernel and independently written ones in other languages. Two of them, con-leche and con-ron, come with a formal proof, relative to set theory, that they cannot accept a false proof. The statement adds that "These checks ran on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof sources."

Andrew Sutherland of MIT has published the [verification files](https://math.mit.edu/~drew/grh/), noting that the "checks run with Claude Code." As of his 12:55 UTC update, all 175,187 modules compiled with zero errors, and the five headline theorems depend only on Lean's three standard axioms. The independent Rust kernel nanoda accepted the full export in 27 minutes 42 seconds on 80 threads, and con-leche accepted 1,468,418 declarations. Lean's own kernels on the export were still running at that point. The compressed proof is 7.7 GB; rechecking it yourself takes roughly 100 GB of RAM and 70 GB of disk. Sutherland's page also sets the scope: "GRH" here means Dirichlet L-functions, not L-functions in general.

On November 20, Ravi Vakil [posted](https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/) a separate proposed proof by Alessandro Maria Masullo, which he described as "all human, with no AI." About eight hours later he wrote that several people had reported problems in it, "identified with the assistance of different AI models."

**Why it matters:** A 17-million-line artifact built with a model was accepted on the strength of a fixed specification checked against it in several independent ways. The setup has three parts worth copying for agent output: a challenge file written apart from the solution, more than one independently written checker, and checking an inert export rather than running the generated sources. It works here because the specification is a formal theorem. Most software has no specification that precise, and writing one is the hard part.
