--- id: "2026-09-27-poincare-conjecture-lean-formalization" url: "https://postcutoff.com/e/2026-09-27-poincare-conjecture-lean-formalization/" as_of: "2026-11-22T13:32:00+01:00" date: "2026-09-27" date_precision: day category: science importance: 4 confidence: medium status: [Result confirmed] verification: Unverified claim sources: 8 editor: human editor human_review: null version: "2026-10-19" --- As of: 2026-11-22 13:32 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-09-27-poincare-conjecture-lean-formalization/ # Perelman's proof of the Poincaré conjecture formalised in Lean, with AI-generated code Full title: Perelman's proof of the Poincaré conjecture formalised in Lean, with AI-generated code (Chow, Qin, Liao, Khaitan) On Sept 27, 2026 Ziyang Qin, Yuan Liao, Ayush Khaitan and Ricci-flow expert Bennett Chow posted a Lean 4 formalization of the Hamilton–Perelman proof of the 3D Poincaré conjecture. Via the Moise smoothing theorem it covers both the smooth and the topological versions, and it proves the statement Mathlib had marked `proof_wanted`. The axiom check shows no `sorry` and no project axioms. The paper says "automated assistance was used for proofs and prose". Secondary reports say about 2.7M of 4.7M Lean lines were produced in the last two weeks with ChatGPT/Codex and Claude. It is the first machine-checked proof of a solved Millennium Prize problem. ## Key facts - arXiv 2609.33842 (math.DG), submitted Sept 27, 2026; code: github.com/qinz1yang/differential-geometry, release v0.1.3 (Lean/Mathlib v4.33.1) - Lean endpoints: exists_isManifold_three (Moise), smoothPoincareConjecture_holds (smooth), DifferentialGeometry.Topology.poincare_conjecture (topological); all four queried closures use only propext, Classical.choice, Quot.sound, with no sorryAx or project-specific axiom (paper) - Proves the topological 3D Poincaré statement marked proof_wanted in Mathlib v4.33.1, checked by applying the endpoints directly (paper) - Route: Ricci flow with surgery and finite-time extinction (Hamilton–Perelman), following Kleiner–Lott and Morgan–Tian - AI disclosure in the paper: 'Automated assistance was used for proofs and prose; the authors retain responsibility'; it reuses parts of Tau Ceti, a Lean library credited to 'J. Tooby-Smith, Codex, and The Tau Ceti contributors' - Secondary report (KuCoin flash citing Chinese outlet Xinzhiyuan): ~4.7M lines in total, ~2.7M in the final two weeks with ChatGPT (Astra), Codex and Claude (Fable 5.1); Qin made 7,477 of 11,000+ commits; ~7 months of work. Unverified against a primary source - Repo README 'AI Disclaimer' (checked Oct 1): 'AI agents assisted with formalizing individual proofs and writing boilerplate'; names 'Generative AI (ChatGPT, Claude, Deepseek, Gemini, GLM, etc.)' without saying which models; 'The high-level architecture is human-designed'; humans checked definitions and core theorem statements; the repo shows 11,175 commits. Line counts and the Astra/Fable 5.1 attribution are still only in secondary reports - Preceded by the same group's 'A Lean Formalization of Hamilton's Three-Manifold Theorem' (arXiv 2608.21502, Aug 2026) - Second, independent formalization (6 Oct 2026): Zhang, Delaval, Chen, Chen, Xu, Liao, Jiang, Liu and Bin Dong (Peking University/BICMR and others), arXiv 2610.08329, code github.com/frenzymath/PoincareConjecture. It follows Morgan–Tian, proves the topological and smooth theorems, passes a full Lean build and Comparator checks against Mathlib-only target statements (also re-checked with the independent Nanoda kernel) - Frenzymath numbers (paper): preparation from May 2026 with three AI engineers and a group of mathematicians; ~90 milestones; main formalization restarted in September and finished in 'slightly more than two weeks' with publicly available commercial models; ~3.2M Lean lines, 2.8M after dependency extraction; cost ~$25,000 (~$20,000 of AI subscriptions, mainly GPT Pro 20× plans, plus ~$5,000 of servers) - Frenzymath models (paper): 'GPT-6-Astra formalized the milestone statements in Lean, and both GPT-6-Astra and Fable5.1 performed multiple rounds of statement review. Proof construction relied primarily on GPT-6-Astra, with occasional consultation of Fable-5.1'; interactive sessions used Codex and Claude Code plus a distributed 'Archon Horizon' mission workflow - Follow-up (9 Oct 2026): the same group, now ten authors, formalized the full geometrization conjecture (arXiv 2610.13897, release v0.1.5) and named its models for the first time: the Humanize builder–reviewer workflow with 'GPT-6-Astra as the reviewer' and builders chosen among GPT-6 Luna, GPT-6 Sol, GPT-6 Astra, Claude Opus 5.5 and DeepSeek V4.1 Flash. That statement is about the geometrization work; the Poincaré paper itself still gives no model names or line counts (see [2026-10-09-geometrization-conjecture-lean-formalization](https://postcutoff.com/e/2026-10-09-geometrization-conjecture-lean-formalization/index.md)) ## What happened Bennett Chow (UCSD), a leading Ricci-flow researcher, and three younger collaborators released a Lean 4 library that formalizes the whole Hamilton–Perelman proof: short-time existence via DeTurck, singularity analysis, surgery and finite-time extinction. With the Moise smoothing theorem it yields the topological Poincaré conjecture. The paper is short (it points to the code) and includes axiom-check commands that anyone can rerun. On 6 October a second, independent team (Frenzymath, led from Peking University by Bin Dong and others) posted its own complete formalization of both the smooth and topological Poincaré theorems, following Morgan–Tian (arXiv 2610.08329). Unlike Chow's group, it documents the AI side in detail: GPT-6 Astra wrote most of the proofs, Claude Fable 5.1 helped review statements, the main phase took just over two weeks, the code is about 2.8 million lines and the project cost about $25,000. Two independent machine-checked proofs now exist. Chow's group's paper discloses AI help in one sentence. The scale (millions of lines in months) and the model names come from secondary reports, and confidence is medium until the authors or the labs confirm them. ## Why it matters After Fermat's Last Theorem (Claude, Sept 2026), this is the second landmark formalization of a famous proof in a month. Perelman's proof needed years of expert refereeing in 2003–06. A kernel-checked version suggests that AI-generated Lean at scale now reaches the hardest parts of geometric analysis. Experts still have to check that the Lean definitions match the mathematics. ## Your AI and this story - GPT-6.2 Astra (training cutoff May 2026): 119 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 89 days after its cutoff - Gemini 4 Hydrogen (training cutoff March 2026): 180 days after its cutoff - Grok 4.7 (training cutoff May 2026): 119 days after its cutoff ## Sources 1. [arXiv: A Lean Formalization of the Hamilton–Perelman Proof of the Three-Dimensional Poincaré Conjecture](https://arxiv.org/abs/2609.33842) (arxiv.org, paper) 2. [GitHub: qinz1yang/differential-geometry](https://github.com/qinz1yang/differential-geometry) (github.com, code) 3. [arXiv: A Lean Formalization of Hamilton's Three-Manifold Theorem (2608.21502)](https://arxiv.org/abs/2608.21502) (arxiv.org, paper) 4. [Frenzymath: Formalizing the Poincaré Conjecture in Lean 4 (separate project)](https://github.com/frenzymath/Poincare-Conjecture) (github.com, code) 5. [arXiv 2610.08329: An AI-Assisted Formalization of the Poincaré Conjecture (Frenzymath/PKU)](https://arxiv.org/abs/2610.08329) (arxiv.org, paper) 6. [GitHub: frenzymath/PoincareConjecture (second, independent formalization)](https://github.com/frenzymath/PoincareConjecture) (github.com, code) 7. [KuCoin News: Poincaré Conjecture proof fully verified by AI in 4.7 million lines of code](https://www.kucoin.com/news/flash/poincar-conjecture-proof-fully-verified-by-ai-in-4-7-million-lines-of-code) (kucoin.com, press) 8. [Bennett Chow: Ricci Flow as a Test for AI (CMSA talk slides, Sept 2025)](https://mathweb.ucsd.edu/~bechow/GML-Talk-CMSA-091625.pdf) (mathweb.ucsd.edu, discussion) ## Changes - 2026-10-19: Fixed ai_system (unquoted commas); status pending → confirmed: two complete Lean formalizations, one Comparator-checked against Mathlib-only statements - 2026-10-12: Added the geometrization follow-up (arXiv 2610.13897), which names the agent workflow and models - 2026-10-07: Science block: kind: formalization (formal proof of a theorem already proved) (not a long-open problem; listed with other AI science results, not in the open-problems count) - 2026-10-07: Added the second, independent Frenzymath/PKU formalization (arXiv 2610.08329; GPT-6 Astra, Fable 5.1, ~$25k, ~2.8M lines) - 2026-10-01: Added the repo README's AI disclaimer (models listed generically; no line counts), so line counts and model names stay unverified - 2026-10-01 (filed): Created ## Related - 2026-10-09: [Thurston's geometrization conjecture formalised in Lean](https://postcutoff.com/e/2026-10-09-geometrization-conjecture-lean-formalization/index.md) - 2026-09-21: [Lean Pool: an AI-maintained archive of Lean formalizations grows past 3 million lines](https://postcutoff.com/e/2026-09-21-lean-pool-ai-maintained-formal-math-archive/index.md) - 2026-09-08: [OpenAI claims a Millennium Prize problem](https://postcutoff.com/e/2026-09-08-openai-navier-stokes-blowup/index.md) - 2026-09-04: [Claude produces the first complete machine-checked proof of Fermat's Last Theorem in Lean, in 11 days](https://postcutoff.com/e/2026-09-04-claude-formalizes-fermats-last-theorem/index.md) - 2026-07-02: [Mistral releases Leanstral 1.5, an open-weights Lean 4 proof agent that saturates miniF2F and solves 587/672 PutnamBench problems](https://postcutoff.com/e/2026-07-02-mistral-leanstral-1-5/index.md) - 2026-03-01: [Math Inc's Gauss formalises Viazovska's sphere-packing proofs in dimensions 8 and 24, fixing errors in the originals](https://postcutoff.com/e/2026-03-01-gauss-sphere-packing-formalization/index.md)