Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · AI and ML research outputs

Anthropic Mathematician Releases 175,000-File Lean Archive Claiming Riemann Hypothesis Proof

Anthropic mathematician Levent Alpöge released a 175,187-file Lean archive claiming a proof of the Riemann Hypothesis developed alongside Claude.

Published 21 Nov 2026· 7 minute read
A university classroom chalkboard covered with mathematical equations and formulas
Photo: LBM1948 / Wikimedia Commons, CC BY-SA 4.0 (cropped). Source · Licence

On November 20, 2026, Anthropic research scientist and Harvard junior fellow Levent Alpöge released a 14-part archive containing 175,187 Lean 4 files claiming a machine-checked proof of the Riemann Hypothesis and Generalized Riemann Hypothesis in an announcement note titled Remark, challenging academic institutions, research offices, and mathematics departments to adapt traditional peer review to massive artificial-intelligence-generated formal codebases.

The 175,187-module formal Lean archive release

Mathematician Levent Alpöge, a member of the technical staff at Anthropic and a junior fellow at the Harvard Society of Fellows, published a four-page technical note titled Remark alongside a 14-part compressed archive containing 175,187 Lean 4 source files. Distributed across compressed tar archives totaling 239.3 megabytes and expanding to more than 61.5 gigabytes of proof data, the release claims an assumption-free formalization of both the classical Riemann Hypothesis and the Generalized Riemann Hypothesis for all Dirichlet L-functions. The announcement marks the first instance of a Millennium Prize Problem being claimed through a fully compiled formal proof assistant codebase rather than a conventional academic manuscript.

Rather than providing an informal mathematical text, Alpöge noted that an AI-generated natural-language exposition of the complete derivation would span more than 250,000 pages of prose, rendering conventional human refereeing impossible. To provide an auditable artifact, Alpöge chose to publish the argument exclusively as machine-checkable code in Lean 4. The underlying derivation was developed in intense collaboration with Anthropic’s Claude model over several months. Alpöge reported guiding the high-level architecture—translating ideas from toy settings into higher-rank automorphic forms—while Claude executed complex algebraic manipulations, trace formulas, and technical lemmas near the conclusion of the proof.

According to Alpöge’s personal account, the final theoretical pieces aligned late on Sunday night, November 15, 2026. Following consultations with Anthropic Chief Executive Officer Dario Amodei and close mathematical colleagues, Alpöge proceeded with the public release. In his note, Alpöge reflected on the collaboration, observing that Claude demonstrated astonishing mathematical proficiency near the finish line, enabling the assembly of deep number-theoretic machinery across automorphic representations and arithmetic geometry.

Technical specifications and independent kernel audits

Following the public release, computational number theorist Andrew Sutherland of the Massachusetts Institute of Technology coordinated an initial technical evaluation on high-performance computing clusters. As documented on Sutherland’s verification dashboard, all 175,187 Lean source modules compiled successfully without errors under Lean version 4.33.1 and Mathlib version 4.33.1. Replay verification through the core Lean kernel (leanchecker) validated all 175,187 compiled modules with zero rejections.

The formal submission defines the primary theorem in Main.lean against Mathlib’s formal definition of the Riemann zeta function, establishing that all non-trivial complex zeros possess real part equal to one-half. Sutherland’s axiom audit confirmed that the headline theorems depend strictly on Lean’s three standard foundational axioms: propositional extensionality (propext), classical choice (Classical.choice), and quotient soundness (Quot.sound). Independent proof checkers, including the Rust-based nanoda kernel and the multi-threaded con-leche checker, processed the exported proof declarations without finding inconsistencies. Through modern formal verification methods, theorem provers establish mechanical correctness directly against primitive inference rules.

Verification Layer Auditing Tool or Kernel Target Scope Audit Outcome
Module Compilation Lean v4.33.1 and Mathlib v4.33.1 175,187 source files Pass (zero compilation errors across all modules)
Core Axiom Dependency Axiom audit in Main.lean 5 headline theorems Pass (relies solely on propext, Classical.choice, and Quot.sound)
Lean Kernel Replay leanchecker kernel replay 175,187 olean files Pass (175,187 of 175,187 modules accepted)
Independent Rust Kernel nanoda (80 threads) 61.6 GB unpacked proof export Accepted (exit code 0 after 27 minutes 42 seconds)
Statement Comparison con-leche (–compare mode) 51,692 challenge declarations Accepted (theorems match challenge specification)
External Type Checking lean4lean kernel (67f63b0c) All compiled modules Pass (175,187 modules confirmed)

Mathematical lineage and AI collaboration strategy

Formulated by Bernhard Riemann in 1859, the Riemann Hypothesis governs the precise distribution of prime numbers and represents one of the most formidable problems in pure mathematics. In 2000, the Clay Mathematics Institute designated it one of seven Millennium Prize Problems, establishing a one-million-dollar bounty for a verified solution. Alpöge’s proof strategy departs from traditional analytic number theory approaches by reviving and extending Stepanov’s method, historically applied to resolve the Riemann hypothesis for curves over finite fields. Because multiplicity is difficult to define over the integers, the approach chains theta lifts across totally definite unitary groups to construct an effective analogue of multiplicity.

The formal argument draws upon diverse mathematical branches, uniting arithmetic algebraic geometry, automorphic forms on higher-rank Lie groups, Iwasawa theory, and spectral analysis. As reported by Quartz, Anthropic evaluated internal publication protocols before clearing the formal announcement, balancing corporate communication policies against the open dissemination expectations of the scientific community. Historically, landmark proofs of major conjectures—such as Andrew Wiles’ proof of Fermat’s Last Theorem published in the Annals of Mathematics—underwent extensive human refereeing spanning several months to validate every conceptual transition.

Community reaction and the distinction between syntax and semantics

The release triggered immediate discussion across mathematical and computer science communities. Analysis published by Unite.AI emphasized that while the sheer scale of the 175,187-module repository illustrates the rapid evolution of AI-assisted formalization, the broader mathematics establishment continues to regard the proof as unverified pending human comprehension. Independent coverage from MLQ.ai and Clarín noted that while Sutherland’s computational tests confirm that the files adhere to formal syntax, formal code verification does not automatically guarantee mathematical understanding.

Senior number theorists have highlighted the vital difference between proof checking and proof comprehension. While kernel replay confirms that no logical contradictions or unauthorized axioms were introduced, mathematicians must scrutinize Mathlib’s definitions to ensure they align faithfully with the classic Riemann zeta function without subtle vacuities. The episode underscores the emerging significance of post-publication peer review, in which distributed networks of domain experts inspect decentralized codebases to extract underlying conceptual ideas.

What the release means for universities and research offices

The arrival of a 175,000-module AI-generated proof presents immediate practical considerations for universities, academic libraries, and research administration offices:

  • High-Performance Compute Infrastructure: Replaying and auditing formal codebases of this magnitude requires robust computational infrastructure. Sutherland’s verification process demanded high-memory cloud instances with more than 100 gigabytes of RAM, hundreds of gigabytes of disk storage, and dozens of processor cores. University research computing offices must anticipate growing demand for formal theorem-proving environments. Modern institutional support depends on resilient reproducibility infrastructure capable of executing containerized formal verification pipelines.
  • Research Data Management and Versioning: Archiving 61 gigabytes of unpacked proof exports and tracking dependencies across exact software releases (such as Lean 4.33.1 and Mathlib 4.33.1) requires strict adherence to institutional data repositories. Research data officers must ensure persistent storage, cryptographic checksum auditing, and artifact preservation for AI-derived mathematical assets.
  • Peer Review and Authorship Guidelines: Institutional review boards, tenure committees, and scholarly publishers face new questions regarding attribution and intellectual contribution in machine-assisted mathematics. Research offices must clarify standards for co-authorship, prompt disclosure, and computational validation when evaluating faculty scholarship incorporating automated theorem generation.
  • Validation Standards for AI Experiments: As AI systems increasingly generate complex scientific workflows, academic departments must implement rigorous standards for reproducible AI experiments, ensuring that all prompts, execution environments, model checkpoints, and intermediate code states are fully documented.

Verification roadmap and upcoming milestones

The validation of Alpöge’s claim now enters a prolonged community review phase governed by standard mathematical conventions and official institutional bylaws. The rules of the Clay Mathematics Institute stipulate that a claimed solution to a Millennium Prize Problem must be published in a qualifying peer-reviewed mathematics journal of worldwide standing and remain open to community examination for at least two years before the scientific advisory board may initiate formal prize deliberation.

Alpöge announced plans to present the conceptual architecture of the proof in a series of academic lectures, workshops, and forthcoming human-readable synthesis papers. The objective of these presentations is to distill the core mathematical mechanisms, isolate key simplifications, and allow number theorists to assimilate the argument without parsing hundreds of thousands of formal code lines. Concurrently, independent formal verification working groups within the Lean community are continuing sandboxed container replays and automated comparator audits to ensure complete reproducibility across all independent kernel architectures.

Sources

Ask CASRAI · full access

Ask about this story

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →