---
type: analysis
slug: research-claude-assisted-riemann-proof-passes-first-lean-kernel-checks-statement-and
title: "Claude-assisted Riemann proof passes first Lean kernel checks; statement and comparator audits still pending"
description: "Anthropic mathematician Levent Alpöge released a Lean-only proof of about 17.4 million lines that he says proves the Riemann Hypothesis. MIT's Andrew Sutherland reports it rechecks with only the three standard axioms. No human-readable proof exists yet, and the tamper-check run has not finished."
published_at: 2026-11-22T00:00:00.000Z
topics: ["claude","anthropic","research"]
source_urls: ["https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/","https://alpo.ge/remark.pdf","https://alpo.ge/","https://github.com/Hertaspace/alpoge-rh-lean-verification","https://github.com/CocoSgt/riemann-hypothesis-lean","https://www.scientificamerican.com/article/the-ai-riemann-hypothesis-proof-is-almost-certainly-true-independent-mathematicians-attest/","https://terrytao.wordpress.com/2026/11/20/unforced-euler-singularity-unforced-navier-stokes-singularity-and-the-generalized-riemann-hypothesis/"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On 20 November 2026, Levent Alpöge posted a short note at alpo.ge/remark.pdf saying that, working with Claude, he had found a proof of the Riemann Hypothesis that extends to the Generalized Riemann Hypothesis (GRH). Alpöge is a number theorist whose website lists his affiliation as Anthropic. He released no conventional paper. The evidence is a Lean 4 project published as a multi-part archive. Scientific American reported the claim the same day and, within hours, reported that independent checkers had rechecked the proof in Lean.

For builders this matters beyond mathematics. If the proof holds, it is the strongest public case yet of a model-assisted result whose correctness rests on a machine checker rather than on human reading. It also shows what that kind of trust requires in practice: kernel rechecks, axiom audits, checks that the formal statement means what it claims, and replays on independent kernels. Each of these steps is a separate claim, and they do not all have the same status yet.

## What we checked

- **Can independent mathematicians inspect and validate the Lean formalization as a proof of the Riemann Hypothesis?** partly confirmed. Andrew Sutherland (MIT) says the proof compiled and was independently rechecked by the Lean kernel with only the three standard axioms, and the headline theorem targets Mathlib's own RiemannHypothesis definition. Kevin Buzzard says a successful comparator run (a check for redefined symbols that also replays the proof on multiple kernels) is still awaited. An unaffiliated GitHub verification lists its axiom, comparator and multi-kernel checks as pending or planned. [6][4][5]
- **Is there a human-readable account of the proof and of what Claude contributed?** unresolved. No conventional written proof has been released. Sutherland says much work remains to understand the details, and Terence Tao expects the digestion process to take some time. Claude's specific contribution is described only by Alpöge himself. [1][7]

## What we found

- Scientific American reports that Alpöge claimed the proof in a statement on his website and says it covers GRH. Andrew Sutherland of MIT called it historic after compiling it, but added that understanding the mathematics is still ahead. We could not extract text from the remark.pdf file itself, so quotes from the note come from Scientific American. [1][2][3]
- The artifact is enormous. An independent verification repository counts 175,187 Lean files and about 17.4 million lines built on Mathlib. A public mirror, which says its src tree matches Alpöge's archive byte for byte, noted that at the time of mirroring no independent build or review had been published. [4][5]
- The headline theorem, riemannHypothesis, targets Mathlib's existing RiemannHypothesis definition. Because that definition comes from outside the archive, the risk that the author defined RH his own way is lower for this one theorem. The four GRH-style theorems use statements defined inside the project, so they need a separate check that the formal statement matches the real one. [5][4]
- Scientific American's follow-up quotes Sutherland: the proof compiled and was independently rechecked by the Lean kernel using only the three standard axioms. He put the chance of a problem at approximately zero. This is his own report. We found no public log of the #print axioms output. [6]
- Not every check is done. According to Scientific American, Kevin Buzzard still awaits a successful comparator run. A comparator run confirms that nothing in the code redefines symbols and that the proof passes several independent kernels. The unaffiliated Hertaspace repository says its own source scan passed. Its axiom check was pending and its comparator and multi-kernel replays were planned, with a full build estimated at about 20 hours. That repository says a Claude agent ran its checks under human direction. [6][4]
- Terence Tao points readers to a Lean Zulip thread as the best place to follow verification. He expects understanding of the proof to take time. We could not load the Zulip thread, so we searched once for a reported flaw, failed comparator run or non-standard axiom. We found none. [7]

## What it means

For builders, the lesson is about how trust is built. A proof of about 17.4 million lines that no human has read is being accepted step by step through cheap, mechanical, reproducible checks: a clean build, #print axioms, a comparator check that statements mean what they claim, and replays on independent kernels. Each step closes a different way the proof could be wrong. Sutherland's report closes the kernel and axiom steps for him. The check that statements are faithful, especially for the project-defined GRH statements, and the multi-kernel replays are still open. If you ship model-generated code or proofs, copy this pipeline. Pin the specification outside the generated artifact, as Alpöge did by targeting Mathlib's RiemannHypothesis definition, and make the audit replayable by outsiders.

Until the comparator run and other public replays report, treat the RH theorem as very likely machine-verified, and treat the GRH theorems and the mathematical understanding as pending. Do not read this as evidence about a shipping Claude product. Anthropic has issued no statement, and Claude's specific contribution is described only by Alpöge.

## Competing readings

- **The Lean proof is valid and its headline theorem is the Riemann Hypothesis as Mathlib defines it. Human understanding lags, but the remaining audits will pass** (likely, about 75%). What would tell it apart: Sutherland reports a kernel recheck using only the three standard axioms [6]; The headline statement is Mathlib's definition, outside the archive [5].
- **The artifact has a hidden flaw, such as a kernel exploit, a redefined symbol or an environment trick, so the comparator or multi-kernel replays fail and RH is not proved** (very unlikely, about 10%). What would tell it apart: The comparator and multi-kernel replays were still pending or planned [4]; No independent build or review existed when the archive was mirrored, and the proof is too large for humans to read [4][5].
- **RH holds, but one or more project-defined GRH statements turn out weaker than or different from the standard GRH, which shrinks the generalized claim** (very unlikely, about 15%). What would tell it apart: The GRH theorems rely on statement definitions local to the project [4]; No human-readable account yet confirms the scope [1][7].

## Base rate

Of 45 earlier research stories about companies, people or products in the New Runtime archive over the last 12 months that are at least 3 months old, 38 (84%) were followed within 3 months by a release story on the same company or product line. In New Runtime's archive, 38 of 45 earlier research stories were followed within 3 months by a release story on the same entity. That pattern concerns product releases and says little about whether a mathematical proof is correct, so it carries little weight here.

## What would change this view

- A published comparator or independent-kernel run (for example Buzzard's or the Hertaspace repo) reports a failure, a non-standard axiom or a redefined symbol (signal): The machine verification does not establish RH, which overturns the main reading.
- Lean experts show that the project-local GRH statements differ from the standard GRH (signal): The generalized claim is narrower than announced, though RH could still stand.
- Anthropic or Alpöge withdraws or qualifies the claim, or a correction is posted at alpo.ge (signal): The author no longer stands behind the result.

## Still open

- Will a public comparator run and replays on independent kernels (nanoda, lean4checker and others) succeed, and will their logs be published?
- Do the project-local GRH statements in Statements.lean match the standard Generalized Riemann Hypothesis for Dirichlet L-functions?
- Which Claude model and setup produced the proof, and will Anthropic describe it?
- When will a human-readable account exist, and will number theorists accept the strategy?

## Sources

1. [Scientific American: Anthropic mathematician claims to prove Riemann hypothesis](https://www.scientificamerican.com/article/anthropic-mathematician-claims-to-prove-riemann-hypothesis-most-famous-problem-in-all-of-math/)
2. [Levent Alpöge, remark (note claiming the proof)](https://alpo.ge/remark.pdf) (primary)
3. [Levent Alpöge personal website](https://alpo.ge/) (primary)
4. [Hertaspace/alpoge-rh-lean-verification](https://github.com/Hertaspace/alpoge-rh-lean-verification)
5. [CocoSgt/riemann-hypothesis-lean mirror](https://github.com/CocoSgt/riemann-hypothesis-lean)
6. [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/)
7. [Terence Tao blog, 20 Nov 2026](https://terrytao.wordpress.com/2026/11/20/unforced-euler-singularity-unforced-navier-stokes-singularity-and-the-generalized-riemann-hypothesis/)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
