Thurston’s geometrization conjecture formalised in Lean
The Hamilton–Perelman proof machine-checked with GPT-6, Claude and DeepSeek agents (Chow’s group)
Event confirmedAwaiting review
Importance: major (4 of 5)The takeaway
Twelve days after the Poincaré conjecture, the same group machine-checked the full geometrization theorem, the result Perelman actually proved. The paper names the models: GPT-6 Astra reviews, and GPT-6 Luna, Sol, Astra, Claude Opus 5.5 and DeepSeek V4.1 Flash build.
Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Lean proof, not yet reviewed
- Importance
- Major (4 of 5)
- Last verified
- 12 October 2026
Your AI and this story
- GPT-6 Astra162 days after its cutoff
- Claude Opus 5.5101 days after its cutoff
- Gemini 3.8 Flash192 days after its cutoff
- Grok 4.7131 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 101 days before it.
Key facts
- Paper: ‘A Lean Formalization of the Hamilton–Perelman Proof of Thurston’s Geometrization Conjecture’, arXiv 2610.13897 [math.DG], 9 Oct 2026; Yuan Liao and Ziyang Qin contributed equally. Affiliations: UC San Diego, Cornell, Princeton AI (Princeton University), Tsinghua University, Amazon AGI, NVIDIA Research, University of Oregon
- Theorem formalized: every connected, closed, oriented smooth three-manifold is a connected sum of prime manifolds; each prime factor contains finitely many disjoint, smoothly embedded incompressible tori, and the interior of every piece cut along them carries a complete metric locally isometric to one of the eight Thurston geometries (finite volume in the hyperbolic case)
- Scope stated by the authors: the statement ‘does not require the cutting tori to form the canonical JSJ family’, does not require the family to be minimal and makes ‘no uniqueness assertion’; the prime convention allows S³ as a factor
- Verification (paper): ‘successful endpoint and complete-library builds, a separate consumer of the expanded public statement, and transitive axiom checks’; the four public declarations yield the same axiom set ‘[propext, Classical.choice, Quot.sound]’ and ‘contain neither sorryAx nor a project-specific axiom’. The authors add that the ‘mathematical interpretation also requires review of the definitions and hypotheses’
- AI role (Section 3, ‘Agent-assisted formalization with Humanize’): the Humanize RLAR workflow (‘Ralph Loop with Agentic Review’) separates a builder agent from an independent reviewer. ‘We use GPT-6-Astra as the reviewer. For the builder, we choose among GPT-6-Luna, GPT-6-Sol, GPT-6-Astra, Claude-Opus-5.5, and DeepSeek-V4.1-Flash based on their observed generation speed in tokens per second (TPS).’ ‘Human authors remain responsible for the intended theorem statements, the mathematical interpretation of the formalization, and the final verification claims.’
- Mathematical route: Hamilton’s Ricci-flow program and Perelman’s completion, following Morgan–Tian and Kleiner–Lott; long-time surgery flow, thick-region hyperbolic pieces, the Kleiner–Lott local-collapse argument for graph-manifold pieces, and Hamilton’s incompressibility argument (with Mostow–Prasad rigidity). ‘A fuller account of the analytic and topological constructions will appear elsewhere.’
- Code: github.com/qinz1yang/differential-geometry, release v0.1.5 (preview), published 9 Oct 2026, Lean and Mathlib v4.35.0-rc3, Apache-2.0; the release ‘adds Thurston’s geometrization conjecture for closed, connected, oriented smooth three-manifolds, and three-dimensional Mostow–Prasad rigidity’. The repository had 74 stars on 12 Oct 2026
- Reused sources (paper): the group’s Poincaré development, parts of Armstrong and Kempe’s DeGiorgi library, Jordan–Schönflies, Classification of Compact Surfaces, Tau Ceti and a canonical-topology development
Show 2 more
- Funding and access: Khaitan acknowledges the DARPA expMath program; Chow thanks OpenAI ‘for providing access to OpenAI models through the ChatGPT for Academic Researchers program’
- Not stated in the paper: line counts, cost or time taken. Not found as of 12 Oct 2026: independent re-checks, expert reactions, press or a Hacker News thread
What happened
Thurston’s geometrization conjecture says every closed three-manifold can be cut into pieces that each carry one of eight standard geometries. It contains the Poincaré conjecture as a special case. Grigori Perelman proved it in 2002–03 by completing Richard Hamilton’s Ricci-flow program, and it took several teams years to write out and check the details.
On 27 Sept 2026 Bennett Chow’s group published a Lean formalization of the Poincaré case. On 9 Oct the group, now ten authors including researchers at Amazon AGI and NVIDIA Research, posted the full geometrization theorem. The Poincaré case needs only finite-time extinction of the flow. Geometrization needs the flow with surgery for all time, the analysis of the thick (hyperbolic) and thin (collapsed) parts, and a proof that the cutting tori are incompressible. The short paper reproduces the Lean statement, explains its exact scope and lists the checks performed.
The Lean code was produced with Humanize, a workflow in which a builder agent writes code in rounds and a separate reviewer agent checks progress against the plan. The paper names the models, which the Poincaré paper had not done: GPT-6 Astra reviewed, and the builder was chosen among GPT-6 Luna, GPT-6 Sol, GPT-6 Astra, Claude Opus 5.5 and DeepSeek V4.1 Flash by generation speed.
Why it matters
It is one of the deepest theorems of modern geometry, and the proof is far longer than anything formalized before 2026. The paper is also a data point on method: the proof ideas are human and decades old, but the formal text was written by commercial models in a builder–reviewer loop. The usual caveat applies. The Lean kernel guarantees the proof of the stated theorem; people still have to confirm that the definitions and the statement say what mathematicians mean. The authors say so themselves, and their statement is weaker than the textbook one in stated ways (no canonical JSJ family, no uniqueness).
What is disputed or not yet verified
| Verification | Formal proof in Lean (kernel-checked, axiom closure reported by the authors); definitions and statement still need independent human review; unrefereed preprint |
|---|
Sources
4 sources from 3 sites. Numbers match the chips in the text.
4 sources: 4 primary
Primary
- arXiv 2610.13897: A Lean Formalization of the Hamilton–Perelman Proof of Thurston’s Geometrization Conjecturearxiv.org, paper
- GitHub: qinz1yang/differential-geometry (release v0.1.5 preview)github.com, code
- Humanize RLAR workflow (humanfia.ai)humanfia.ai, docs
- arXiv 2609.33842: the group’s Poincaré conjecture formalizationarxiv.org, paper
Changes
- Filed from the arXiv PDF and the GitHub release notes