Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI

0
2



The Lean and Mathlib communities certified on November 21, 2026 that Levent Alpöge’s 17-million-line Lean library correctly states and proves the Riemann Hypothesis, following a day of compilation, statement-matching, and proof-kernel checks. Their joint statement was endorsed by the Lean FRO and the Mathlib Initiative.

The certification follows Alpöge’s announcement on November 20, 2026, in which he said he discovered and formalized the proof while working with Claude. The release contained only the Lean formalization; Alpöge viewed a conventional write-up at this stage as a huge, unreadable document and planned to understand, simplify, and explain the mathematics properly.

How the Proof Was Checked

The codebase depends directly only on Mathlib, the community’s library of formalized mathematics. The statement being proved was already part of Mathlib and had been vetted by experts in Lean and number theory, addressing the requirement that the formal statement express the intended hypothesis. Dozens of people reported compiling the library without errors or build failures and reproducing the same raw proof export.

To check the theorem’s identity, a community-added challenge file isolated the Riemann Hypothesis statement and imported only Mathlib. Comparator and con-leche’s comparison mode independently confirmed that Alpöge’s theorem matched that statement exactly. The raw proof export was checked with Lean’s official kernel, con-leche, con-ron, and additional third-party checkers written in different programming languages.

Those checks ran on an inert export, beyond the reach of hypothetical malicious code in the Lean sources. Con-leche and con-ron also have formal proofs, relative to set theory, that they cannot accept a false proof, addressing concerns that a flaw in Lean’s logic could be exploited.

What the Verification Establishes

Andrew Sutherland’s MIT verification record, last updated at 12:55 UTC on November 21, 2026, reports that all 175,187 modules compiled with zero errors using Lean and Mathlib version 4.33.1; his checks were run with Claude Code. An axiom audit found that the five headline theorems depended only on propext, Classical.choice, and Quot.sound, with no axioms declared outside Lean core. Nanoda, an independent Rust kernel, and con-leche each accepted the full exported proof, lean4lean accepted every compiled module, and con-leche’s comparison mode matched every challenge statement.

At that page’s cutoff, the Lean default and paranoid kernel checks on the export and the full lake comparator run remained in progress. Older bundled con-leche and con-ron binaries exited without a verdict because they did not support one reflexive inductive type, not because they rejected the proof. Sutherland also disclosed that the build ran on unsandboxed cloud machines and that its export hash had not been compared with one from a sandboxed lake build; a passing result still relies on the kernels used and on Mathlib’s definitions representing their intended mathematics.

The central theorem uses Mathlib’s formal definition: every zero of the Riemann zeta function other than the trivial zeros at negative even integers must have real part one-half. The separate condition s ≠ 1 appears because Mathlib assigns the zeta function a conventional value at its pole. The submission also verifies Generalized Riemann Hypothesis statements for Dirichlet L-functions across all moduli and characters; Sutherland expressly excludes Dedekind zeta functions, Hecke L-functions, and general automorphic L-functions from the checked scope.

Claude’s Role in the Discovery

Alpöge’s account assigns Claude a role in discovery as well as formalization: he guided the work from a toy setting into an analogous target situation before Claude outpaced him near the finish. In one example, Claude suggested that a von Mangoldt sum of length X should use a chain of theta lifts of length X, a technical choice Alpöge highlights in his explanation of the argument.

Alpöge describes the current argument as extremely long, expects substantial simplification, and is still working to understand and explain it. His account describes human guidance and Claude’s technical contributions, with both formalizing the arguments without extra assumptions. The certification concerns the formal artifact; that account supplies the basis for describing Claude’s role, rather than demonstrating an autonomous solution.

The certification resolves the formal-checking question without resolving the explanatory one. Whether the proof’s new ideas can be simplified and absorbed by mathematicians remains the central open question around the verified Lean library.



Source link

LEAVE A REPLY

Please enter your comment!
Please enter your name here