The Riemann proof is verified – what 'checked' must mean for AI work in your company
A Lean library claiming a proof of the Riemann hypothesis was certified by the Lean and Mathlib communities one day after release. The transferable part for enterprises is the verification chain: statement check, reproducible build, independent checkers.
CogLake Redaktion · 22 November 2026 · 5 minThis article was produced by CogLake's AI editorial team in an automated workflow and checked before publication. It reports only what the cited sources say; we ran no tests of our own.
The short version. On 20 November 2026 an Anthropic mathematician published a Lean library claiming a proof of the Riemann hypothesis, developed with Claude. About one day later the Lean and Mathlib communities certified both statement and proof. The transferable lesson for enterprises is not the mathematics but the certification chain: a machine-generated artifact became discussable because four specific checks were run in public – and because the people running them were not the producer.
What was published
Levent Alpöge, a mathematician employed by Anthropic, posted a short note titled simply "The Riemann Hypothesis" on his personal website on 20 November 2026. In it he writes that he and Claude "discovered a proof", that the method extends to the Generalized Riemann Hypothesis, and that the arguments were formalised "without assumptions". According to The Guardian, he describes roughly six weeks of work with the model.
He published no prose proof. What he released is a formal artifact: a Lean library of 17 million lines, split into 14 archives, because a readable version would run past 250,000 pages – his own figure. Alpöge also states plainly that he has not fully mastered large parts of the mathematics around the argument, and that he expects it to be shortened soon. He calls the timing, one day after Anthropic's IPO, "horrible", and says he asked for permission to announce rather than wait.
What "verified" meant, concretely
On 21 November 2026 at 13:55, the Lean and Mathlib communities published a joint statement endorsed by the Lean FRO and the Mathlib Initiative: they have "full confidence that Alpöge's proof of the Riemann Hypothesis is correct". The statement does not rest on trust. It certifies four separate things:
- The statement is the intended one. The theorem proved is Mathlib's own
RiemannHypothesisdefinition, previously vetted by Lean and number-theory experts. - The codebase compiles. Dozens of people independently reproduced the same raw proof export without build failures.
- It proves the right statement. Two independent tools (comparator and con-leche) compared the theorem in Alpöge's library against the Mathlib statement and found them identical.
- The proof is valid. The raw export was checked by Lean's official kernel plus several independent checkers, two of which ship with a formal proof that they cannot accept a false proof. The checks ran on an inert export, out of reach of potentially malicious source code.
That is close to the procedure Kevin Buzzard had spelled out ten days earlier. In "…and the proof is checked in Lean" (12 November 2026) he explains why the phrase alone proves little, and applies four questions to AI-generated repositories. His example: of 722 papers in OpenAI's October "theorem dump", about 45 per cent came with Lean code – but in 38 of the 290 Lean-accompanied theorems the code does not even state the main theorem of the paper, and three papers were retracted over a sign error.
| Check | Question in mathematics | Question in enterprise AI work |
|---|---|---|
| Statement fidelity | Does the Lean statement match the informal theorem? | Does the formal definition match the business question? |
| Reproducible build | Does the code compile for people other than the author? | Can a second team rerun the analysis from the same inputs? |
| Right theorem | Is the intended statement the one proved? | Does the result answer the question asked, not a convenient proxy? |
| Independent validation | Do external checkers accept the proof? | Did someone outside the producing team test the claim? |
The clock is a governance problem
Hours after Alpöge's note, Stanford mathematician and AMS president Ravi Vakil posted a "hurried announcement" of an independent draft of the Generalized Riemann Hypothesis by Alessandro Masullo – all human, no AI, reviewed for months by Stanford's Brian Conrad. Within about eight hours Vakil added an update: several readers had reported problems in the draft, some found with the help of other AI models. "All issues need to be resolved to the satisfaction of the community before this proof would be accepted as true," he wrote.
Two mechanisms collide here. One is the incentive to publish on a company's news cycle; the other is verification that reliably takes longer than an announcement. Vakil's own update is the useful part: problems surfaced within hours not because the draft was absurd, but because other people started checking it immediately.
What to copy
- Publish the artifact, not the story. Alpöge shipped the machine-checkable form and skipped the unreadable prose. In enterprise terms: ship the query, the data extract and the versioned result, not only the slide.
- Define the statement before the answer. Most disputes about AI output are disputes about what was actually asked. Write the definition down and let someone approve it.
- Separate producer and checker. The certification came from communities with no stake in Anthropic's share price.
- Keep a readable layer. Even here a human summary is promised, because 250,000 pages are not knowledge transfer. A result nobody can explain is not institutional memory.
- Time-box the release gate. If publication pressure can override verification, verification will lose – as the Stanford draft shows within eight hours.
Limits
Mathematics is the easy case: a formal proof is decidable, and a kernel either accepts it or does not. Claims about markets, contracts, incidents or policy are not decidable that way, and nothing in this case is evidence about model quality, cost or return. It is not investment commentary either. What the case does demonstrate is that "the AI produced it" and "it has been checked" are two different claims, and that the second one has a procedure.
Sources
- Levent Alpöge, "The Riemann Hypothesis", note, 20 November 2026 – https://alpo.ge/remark.pdf
- Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative, "Lean verification of the Riemann Hypothesis", 21 November 2026, 13:55 – https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
- Kevin Buzzard, "…and the proof is checked in Lean", Xena project, 12 November 2026 – https://xenaproject.wordpress.com/2026/11/12/and-the-proof-is-checked-in-lean/
- The Guardian, "Two groups verify mathematician's claim that Claude helped prove mathematics problem", 21 November 2026 – https://www.theguardian.com/technology/2026/nov/21/anthropic-riemann-hypothesis-solved-claude-ai
- THE DECODER, Manuel Uth, "Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft", 21 November 2026 – https://the-decoder.com/hours-after-anthropic-claims-a-claude-proof-of-the-riemann-hypothesis-stanford-releases-an-all-human-competing-draft/