Lean’s maintainers certify Alpöge’s Riemann proof. Sutherland’s page still has two checks without a verdict
The Lean FRO and the Mathlib Initiative, in a statement dated 13:55 UTC on 21 November, say they certify Levent Alpöge’s proof. Andrew Sutherland’s MIT page, updated at 12:55 UTC, says 175,187 modules compiled and that con-leche matched all five theorems to Mathlib’s statements. It also says the Lean kernels on the exported proof, and a full comparator run, were still going. The page limits the checked “GRH” to Dirichlet L-functions.
Feature6 minIn-depth
James Okonkwo6 min read
- Anthropic
- Claude
- Riemann hypothesis
- Lean

Photo: Dustin Mudry / Shutterstock, via Scientific AmericanSource page
The Lean FRO and the Mathlib Initiative said on Saturday that Levent Alpöge’s proof of the Riemann hypothesis is correct. Their statement, posted on the Lean community blog and dated 13:55 UTC on 21 November, says they have full confidence in the proof, congratulate him, and certify both the statement and the proof. The library, they write, is 17 million lines and depends directly only on Mathlib. An hour earlier, Andrew Sutherland’s page at MIT, last updated at 12:55 UTC, still had two checks marked running and no verdict on either. The page says the checks were run with Claude Code. It also says the results on it are as of that update, and that the page will change as the remaining checks finish.
Alpöge, a mathematician at Anthropic, posted the claim on his own site on 20 November. The note says he believes that working with Claude he has a proof, and that the method resolves the generalized Riemann hypothesis. Clara Moskowitz’s Scientific American story, dated 20 November and labeled as developing, quotes Sutherland: “The Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms,” and “Honestly the chance that anything is fishy here is approximately zero.” The same story quotes Kevin Buzzard of Imperial College London, writing on a Lean community chat, that a successful comparator run was still awaited, to check that the code does not redefine symbols in the statement and, in paranoid mode, that it compiles under several independent kernels. The editor’s note on that story is dated 20 November.
Five theorems, and a narrower generalized claim
Sutherland’s page says the published Lean source compiles without errors. The row for that check says all 175,187 modules compiled on Lean v4.33.1 and Mathlib v4.33.1, with zero errors, and that the run finished at 16:43:21 UTC. The page does not put a calendar date on that clock time. A separate row says a replay of every module through the leanchecker kernel passed, 175,187 of 175,187, with zero rejections, finishing at about 18:17 UTC. Printing the axioms of the five theorems in Main.lean returned only propext, Classical.choice and Quot.sound. The page says no axioms are declared outside Lean’s core: 15 axioms in the environment, all in Init. Mathlib’s compiled files matched the official cache byte for byte, 8,322 Mathlib files and 378 dependency files.
The five theorems are the ones in the submission’s Main.lean. The page prints them as they appear in a challenge file that imports only Mathlib. The first is Mathlib’s own RiemannHypothesis: every zero of the zeta function, other than the trivial zeros at the negative even integers, has real part 1/2, with the point s = 1 excluded because Mathlib gives zeta a conventional value at its pole. The second is the same statement in the submission’s own words. Lean treats the two as equal by definition. The other three are about Dirichlet L-functions. One says L(χ, s) is not zero when the real part of s is greater than 1/2, for every modulus and every character, except the pole of the trivial character at s = 1. One says that a zero inside the critical strip has real part 1/2. One says the same for primitive characters, which the page calls the usual statement of the generalized Riemann hypothesis for Dirichlet L-functions.
The page then draws a line around that list. “GRH” in these checks means Dirichlet L-functions, all moduli and all characters. Dedekind zeta functions, Hecke L-functions and automorphic L-functions in general are not among the statements that were checked. That is a smaller set than the sentence in Alpöge’s note, which says the method resolves the generalized Riemann hypothesis and that he expects, with more time, the extended Riemann hypothesis as well.
A certificate at 13:55, and two rows still marked running
The Lean statement gives four reasons, pointing at a note by Kevin Buzzard on what it takes for a Lean proof to count as checked. The statement of the Riemann hypothesis is Mathlib’s, and has been vetted by people who work in Lean and in number theory. The library compiles: the statement says dozens of people have reported a correct compilation and the same raw proof export, with no errors or build failures. The code proves that statement: a file that isolates Mathlib’s Riemann hypothesis, and imports only Mathlib, was compared with Alpöge’s theorem by two tools, comparator and con-leche in compare mode, and both confirmed a match. The proof is valid: the exported proof was checked by Lean’s official kernel, by con-leche and con-ron, and by other checkers written in other languages. The statement says con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, and that the checks ran on an inert export, out of reach of any malicious code in the Lean sources. It points to Sutherland’s page for the overview, and it says the community still has to explain the new ideas.
Sutherland’s table, timestamped an hour before that post, does not read as a finished list. nanoda, an independent kernel written in Rust, accepted the full export, exit 0 at 19:05:45 UTC, in 27 minutes and 42 seconds on 80 threads. con-leche at master, commit cd2eceab, accepted 1,468,418 declarations in verified mode, exit 0 at 21:19:04 UTC. Its compare mode, on a branch at 5dc36dc1, accepted all five theorems as “justified by the solution,” exit 0 at 21:19:12 UTC. lean4lean, at 67f63b0c and built with Lean v4.33.1, passed every compiled module, finishing at about 21:55 UTC: 175,182 on the default settings, and five more after a rerun with the step limit raised, because those five had hit a deterministic timeout on decide computations. The same table says the copies of con-leche and con-ron bundled with Lean v4.36.0 declined, exit 2, and that this was not a rejection: those older binaries do not support one reflexive inductive type in the proof. con-leche at master does. Two rows were still open at the 12:55 update. The Lean default kernel and the Lean paranoid kernel, run on the export and single-threaded, had no verdict. A lake comparator run on Lean v4.36.0 was still going. The page says that in one run the statement comparison and the axiom check had already passed, and that the kernels in that run were in progress.
The Saturday statement says Lean’s official kernel has checked the export. The page it cites as the overview still, at the update an hour earlier, had that export replay marked running, and a module-by-module leanchecker replay marked passed. The two documents do not use the same sentence for the same row. Neither says the Clay Mathematics Institute has accepted the proof.
Two inequalities with no zeta in the statement
As a check that does not use Mathlib’s definition of zeta, Sutherland’s page says the proof was combined with an independent formalization of Robin’s criterion and Lagarias’s criterion, kimihiro64/Robin1984 at commit acab1a3, on the same Lean and Mathlib versions. Importing both produces robin_inequality and lagarias_inequality, statements about the divisor sum with no hypothesis and no mention of zeta. The page says those two compile, and that their axioms are again only the three standard ones. con-leche at master accepted the combined export, 1,472,825 declarations, exit 0 at 02:29:56 UTC. Compare mode accepted both theorems. nanoda accepted the combined export at 02:56:46 UTC. The lake comparator run on that combined export was still going at the update: the statement comparison and the axiom check had passed, and the Lean kernel stages, single-threaded on a 62-gigabyte export, had not. The page says those statements still rely on Mathlib’s definitions of the divisor sum, Euler’s constant, harmonic numbers, exp and log, and on the Robin1984 proof.
The caveats on the page are about the build, not about a second version of the theorem. The compile ran on cloud machines without a sandbox. Each file was compiled by calling lean directly, with the inputs lake would have supplied, not through lake build for the full set. The hashes of the compiled files depend on the build directory, because Lean v4.33.1 stores the absolute path of a source file inside linter messages, and 98,219 of the 175,187 compiled files contain the path /home/ubuntu/grh/src/lean. The exported proof is the file the independent kernels checked. A pass, the page says, shows that the formal statements, as written in Lean against Mathlib’s definitions, follow from the three standard axioms according to the kernels that were used.


