# Claude's claimed Riemann proof compiles: what the Lean checks cover and what they leave open

Published: 2026-11-21

Levent Alpöge, a mathematician at Anthropic, posted a four-page note saying he believes he has, working with Claude, discovered a proof of the Riemann Hypothesis and its generalization, and released only Lean code: about 17.4 million lines in 175,187 files. He posted it on Friday, November 20, 2026, and it is a claim, not an accepted proof. As of Saturday November 21 at 17:05 GMT+7, Andrew Sutherland of MIT reported that all 175,187 modules compile with Lean v4.33.1 and Mathlib, that the five headline theorems depend only on the three standard axioms (propext, Classical.choice, Quot.sound), that the independently written kernels nanoda and con-leche accept the full 61.6 GB exported proof (27 min 42 s and 3 min 15 s), and that con-leche's compare mode, an independent implementation of Comparator's statement matching, accepts all five theorems. The official \`lake comparator\` run and the Lean default and paranoid kernel replays on the full proof were still running. Kevin Buzzard's four questions for AI-generated Lean (does the statement match, does it compile, does the code prove the stated theorem, does it exploit a kernel bug) frame what each check covers. Sutherland also checked Robin's inequality, a statement about divisor sums with no zeta function, which independent checkers accepted. A second claim, Alessandro Masullo's human-written draft proof of the generalized hypothesis announced by Ravi Vakil, drew reports of problems found with AI help; Vakil makes no claim as to its correctness and Brian Conrad told Scientific American it is undergoing major revision. Terence Tao wrote that digesting the Lean-only proof will take some time. Anthropic's shares closed at $88.91 on Friday, about 8.5 percent below Thursday's close; The Guardian attributes the drop to a broad tech sell-off and draws no link to the proof.

Canonical: https://thedailydiff.dev/video/2026-11-21-lean-riemann-proof/

## What this video covers

- A Lean proof compiling is a real check: Andrew Sutherland (MIT) reports 175,187 of 175,187 modules compile with zero errors, and the five headline theorems use only propext, Classical.choice and Quot.sound (page last updated 2026-11-21 10:05 UTC).
- Independent kernels agree: nanoda (Rust) accepted the 61.6 GB export in 27 min 42 s on 80 threads and con-leche in 3 min 15 s on 64 threads; Joachim Breitner of the Lean community called the result very thoroughly checked.
- Kevin Buzzard's four questions (Xena Project post, November 12) are the checklist: does the Lean statement match the informal one, does the code compile, does the code prove the theorem as stated (lake comparator), does it exploit a kernel bug (lake comparator --paranoid or several independent kernels).
- Status as of Nov 21 17:05 GMT+7: the official lake comparator run on the full proof and the Lean default and paranoid kernel replays were still running; its statement comparison and axiom check had passed in one run, and con-leche --compare accepted all five theorems.
- The statement is Mathlib's own RiemannHypothesis definition; Sutherland also tested Robin's inequality, which never mentions zeta, and says that effectively rules out holes in the definitions behind the Mathlib statement. His page also states the result relies on the kernels and on Mathlib's definitions being as intended.
- A compiling Lean file does not explain the mathematics: Alpöge says a prose version he generated hit 250,000+ pages and posted only Lean; Terence Tao expects digestion to take time; Scott Aaronson says mathematicians will be satisfied when they can explain it themselves.
- A rival human-written draft (Alessandro Masullo, announced by Ravi Vakil) drew AI-assisted reports of problems; Vakil makes no claim as to its correctness and Brian Conrad says it is undergoing major revision.
- For builders: read the spec before the code, ask what compares the proof to the statement and who wrote the statement, run a second checker written by someone else, and note that compiled-file hashes can depend on the build directory (98,219 of 175,187 files embed a path).
- Visual credits: portrait of Terence Tao at IPAM's Fireside Chat 2026 by the Institute for Pure &amp; Applied Mathematics, CC BY 4.0, cropped with the background removed; portrait of Dario Amodei at TechCrunch Disrupt 2023 by TechCrunch, CC BY 2.0, background removed; each identifies only that person or Anthropic and implies no endorsement, quote or reaction; the screen recording and excerpts of math.mit.edu, Lean Zulip, alpo.ge, Ravi Vakil's blog and the other cited pages are credited to their own pages; a two-second silent quotation of the "Double checking" corgi GIF via GIPHY, whose owner retains the rights and which implies nobody in this story reacted; the Is This A Pigeon meme template via imgflip.

## Questions and answers

### What does it mean that the Lean proof of the Riemann Hypothesis compiles?

Andrew Sutherland of MIT reports that all 175,187 modules of Levent Alpöge's Lean archive compile with Lean v4.33.1 and Mathlib with zero errors, and that the five headline theorems depend only on the three standard axioms propext, Classical.choice and Quot.sound. Compiling shows the Lean statements follow from those axioms according to the checkers used. It does not explain the mathematics, which mathematicians have not yet digested.

### What are Kevin Buzzard's four questions for checking an AI-generated Lean proof?

In a November 12 post, Kevin Buzzard lists four questions. Does the Lean statement match the informal statement? Does the code compile? Does the code prove the theorem as stated, which the lake comparator tool checks? Does it exploit a bug in Lean's official kernel, which several independent kernels or comparator in paranoid mode address? He writes that AI-generated code needs these tests before any correctness claim.

### What does Lean's comparator check, and where did its run on the Riemann proof stand?

Comparator checks that the theorem stated in a trusted file and the theorem proved through untrusted imports are identical, and that only the standard axioms are used (Kevin Buzzard). Andrew Sutherland's MIT page, updated November 21 at 10:05 UTC, listed the official run on the full proof as still running, with its statement and axiom stages passed in one run, while con-leche's compare mode accepted all five theorems.

### Did independent Lean checkers agree on the Alpöge Riemann Hypothesis proof?

According to Andrew Sutherland's MIT page, the independently written kernels nanoda and con-leche accepted the full 61.6 GB exported proof, taking 27 minutes 42 seconds and 3 minutes 15 seconds, and lean4lean accepted every compiled module. Lean developer Joachim Breitner wrote that the proof is now very thoroughly checked. The official comparator run and Lean's own kernel replays were still in progress at that update.

### What was The Daily Diff's verdict on the Lean Riemann Hypothesis proof claim?

The Daily Diff's verdict was NEEDS REVIEW. The compile is real and independent checkers agree, but the official comparator run was unfinished at the last update and mathematicians have not yet digested the argument. Terence Tao expects digestion to take time, and Scott Aaronson says mathematicians will be satisfied when they can explain the ideas themselves. A rival human draft by Alessandro Masullo is under major revision.

## Chapters

- 0:00 What did strangers really confirm?
- 0:29 Who claims what, and in what form?
- 0:55 What are the four questions?
- 1:20 Does it compile?
- 1:36 Could Lean's own kernel be fooled?
- 2:06 Is it the theorem we think it is?
- 2:34 What about the human draft?
- 3:00 What can't a checker tell you?
- 3:25 What does a builder do with this?
- 3:50 Where did the missing check stand?
- 4:36 What's today's verdict?

## Transcript

### What did strangers really confirm?

0:00 Most channels are telling you Claude proved the Riemann Hypothesis. What strangers have confirmed is that seventeen million lines of Lean compile, and that one checker accepted the whole file in under four minutes. In this video, what did that checker actually check? What can a checker never tell you? And what should a builder copy from it? One more check was still running on Saturday afternoon, the one Kevin Buzzard said he was waiting for.

0:23 I'll tell you where it stood before the verdict. It's Saturday, November twenty-first, and this is The Daily Diff.

### Who claims what, and in what form?

0:29 Levent Alpöge is a mathematician at Anthropic. On Friday he posted a four page note saying he believes he has, working with Claude, discovered a proof of the Riemann Hypothesis. He released no paper, only Lean, a language where a program checks every step of a proof. The IPO I covered yesterday is the backdrop. Anthropic's shares fell more than eight percent on Friday, in a sell-off The Guardian does not tie to the proof,

0:53 and neither do I. The archive holds one hundred seventy-five thousand files.

### What are the four questions?

0:58 In September I told you Claude's Fermat proof was thirteen million lines, and Kevin Buzzard ran out of memory trying to compile it. Nine days ago he wrote a four question checklist for AI written Lean. One, does the Lean statement mean the real theorem? Two, does the code compile? Three, does the proof prove that statement and nothing sneakier? Four, does it exploit a bug in Lean's own kernel? Question two is the easy one.

### Does it compile?

1:21 Andrew Sutherland at MIT compiled all of it with zero errors, and the five headline theorems rest only on the three axioms Lean proofs normally use. Buzzard says that answer is yes. Running the official check yourself needs about a hundred gigabytes of memory.

### Could Lean's own kernel be fooled?

1:36 Question four is the paranoid one. Scott Aaronson points out that Lean's kernel has had bugs that allowed proofs of zero equals one. So other people wrote their own checkers, in other languages, and fed them the exported proof. Nanoda, written in Rust, accepted it in under half an hour. Con-leche, which its authors back with a consistency proof, took about three minutes.

1:56 Joachim Breitner of the Lean community called it very, very thoroughly checked. Aaronson adds that no AI result with a Lean certificate has been retracted so far. Questions one and three are the hard ones.

### Is it the theorem we think it is?

2:07 The statement is Mathlib's own definition of the Riemann Hypothesis, and its comment says a proof is worth a million dollars. But seventeen million untrusted lines sit between that statement and the proof, and any of them could quietly redefine a symbol and prove something easier. Comparator, the tool I explained on November tenth, exists for exactly that. It compares the trusted statement with the theorem the proof actually delivers, and it checks the axioms.

2:33 A second claim arrived the same morning.

### What about the human draft?

2:35 Ravi Vakil, president of the American Mathematical Society, posted a draft general Riemann proof by Alessandro Masullo. It was written with no AI, and Brian Conrad had been reviewing it for months. Eight hours later Vakil added that people using AI models had found problems, and that he makes no claim it is correct. Conrad told Scientific American it is undergoing major revision. Vakil also wrote that he really hopes it works. A checker tells you whether the steps follow.

### What can't a checker tell you?

3:02 It cannot tell you why the thing works. Alpöge's note says Claude flew, and danced, and sparkled near the finish line. He says a prose version he generated hit two hundred fifty thousand pages, which he called slop, so he posted only the Lean. Terence Tao wrote that digesting it will take some time, and Scott Aaronson says mathematicians will be satisfied when they can explain it themselves. Nothing in your stack changes tonight,

### What does a builder do with this?

3:26 since Scientific American says the result has no immediate real-world applications. For builders, three habits. Read the spec before the code. Ask what compares the proof to the statement, and who wrote the statement. And run a second checker written by someone else. Even build hygiene shows up, because Sutherland notes the compiled files' hashes change with the folder you build in, since Lean can store the source path inside

3:49 them. So where did that check stand Saturday afternoon?

### Where did the missing check stand?

3:52 In developer terms, CI was green except the one job everybody was watching. The official comparator run on the full proof was still going. Its statement comparison and axiom check had already passed in one run, and the slow part replays the proof in Lean's own kernels, one thread at a time. Meanwhile con-leche's compare mode, an independent copy of the same statement matching, accepted all five theorems. Sutherland also tried a statement with no zeta in it at all,

4:18 Robin's inequality, an old equivalent of the hypothesis about sums of divisors. The independent checkers accept that one too, and Sutherland says that effectively rules out holes in the definitions behind the Mathlib statement. If you'd rather read this than hear me say it, the diff lands in your inbox every morning, free at the daily diff dot dev, link below.

### What's today's verdict?

4:36 So today's verdict is NEEDS REVIEW. I'd ship the checking and hold the conclusion. The compile is real, strangers' checkers agree, and the official run is unfinished, while mathematicians have not yet digested the argument. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

## Sources

- [Levent Alpöge, note announcing a Lean proof (PDF)](https://alpo.ge/remark.pdf) — Levent Alpöge
- [RH Lean proof: verification files](https://math.mit.edu/~drew/grh/) — Andrew Sutherland, MIT
- […and the proof is checked in Lean](https://xenaproject.wordpress.com/2026/11/12/and-the-proof-is-checked-in-lean/) — Kevin Buzzard, Xena Project
- [Discussion: Alpöge Proves the Riemann Hypothesis](https://leanprover.zulipchat.com/) — Lean Zulip (#general)
- [Mathlib: RiemannZeta.lean (RiemannHypothesis)](https://github.com/leanprover-community/mathlib4/blob/master/Mathlib/NumberTheory/LSeries/RiemannZeta.lean) — leanprover-community
- [Hurried Announcement: Alessandro Masullo GRH](https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/) — Ravi Vakil
- [Unforced Euler singularity, unforced Navier-Stokes singularity, and the Generalized Riemann Hypothesis](https://terrytao.wordpress.com/2026/11/20/unforced-euler-singularity-unforced-navier-stokes-singularity-and-the-generalized-riemann-hypothesis/) — Terence Tao
- [The stochastic parrots have apparently proved the Riemann Hypothesis](https://scottaaronson.blog/?p=10344) — Scott Aaronson
- [Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math](https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/) — Scientific American
- [The AI Riemann hypothesis proof is almost certainly true, independent mathematicians attest](https://www.scientificamerican.com/article/the-ai-riemann-hypothesis-proof-is-almost-certainly-true-independent-mathematicians-attest/) — Scientific American
- [Dueling claims of Riemann hypothesis proofs emerge](https://www.scientificamerican.com/article/dueling-claims-of-riemann-hypothesis-proofs-emerge/) — Scientific American
- [Nasdaq erases all 2026 gains as AI sell-off dampens Anthropic's second day of trading](https://www.theguardian.com/business/2026/nov/20/global-stock-markets-slump-anthropic-ipo-ai-sell-off) — The Guardian
- [Levent Alpöge claims proof of the Generalized Riemann Hypothesis \[pdf\]](https://news.ycombinator.com/item?id=50574968) — Hacker News
