Lean FRO

Velvet 2.0 is out: now based on Lean's most recent verification machinery, easier to set up, and 10x faster. New: exception specs, ghost state, named proof goals, lots of case studies from Dijkstra to lazy segment trees. And a shiny new webpage: https://velvetprover.dev/

Our research team at the University of Pennsylvania is looking for participants for a study to evaluate the usability of a new interactive tool for writing and debugging automation-heavy Lean proofs! The study will consist of a ~1 hour recorded Zoom call, during which you will write a few proofs and answer some questions about the experience. Compensation is $30. If you are interested in participating , please fill out this form: https://lnkd.in/ejHHUpE2 Send any questions to Cassia Torczon, ctorczon@upenn.edu

We're grateful to XTX Markets for their support of Lean FRO.

We’re excited to launch Lean Kernel Challenge Stage 1, co-organized by Lean FRO and the SAIR Foundation! The Lean Kernel Challenge is a multi-stage competition to improve the performance of verified computation in the Lean 4 kernel that the whole community can benefit from. Stage 1 begins with eight problems across algebra, number theory, combinatorics, cryptography, and discrete mathematics. Develop faster algorithms, prove their correctness in Lean, and compete on separate leaderboards ranked by kernel computation instruction count. Our organizing committee includes Joachim Breitner, Leonardo de Moura, Kim Morrison, and Terence Tao. Submission deadline: November 20, 2026, 23:59 AoE (UTC−12). Explore the competition and participate: https://lnkd.in/g7vd6vFG

𝐋𝐞𝐚𝐧 𝟒.𝟑𝟒.𝟎 𝐢𝐬 𝐥𝐢𝐯𝐞! This release brings 159 changes, including kernel soundness hardening, significant 𝚋𝚟_𝚍𝚎𝚌𝚒𝚍𝚎 improvements, and extended linter capabilities. Notable improvements include: 🔧 Three kernel soundness vulnerabilities fixed, plus a round of additional defensive checks. All three required deliberately constructed inputs and matter when checking proofs from an untrusted source, not in ordinary development. ⚡ 𝚋𝚟_𝚍𝚎𝚌𝚒𝚍𝚎 is up to 6x faster, with new integration in 𝚐𝚛𝚒𝚗𝚍 interactive mode and improvements to bit-vector preprocessing. 🔍 Stateful linters now persist across elaboration phases with a two-phase architecture. New code action support enables editor-integrated fixes for deprecation warnings. 🛠️ Lake improvements across linting, caching, and error reporting. 𝐅𝐮𝐥𝐥 𝐫𝐞𝐥𝐞𝐚𝐬𝐞 𝐧𝐨𝐭𝐞𝐬: https://lnkd.in/ggZAtUG6 #LeanLang #LeanProver #ProofAssistant #OpenSource #FormalVerification

I'm co-organizing "AI for Verifiable Coding," a NeurIPS 2026 workshop on AI and formal verification in both directions: AI that writes specs, proofs, and verified code; and verification that keeps AI systems in check. 📍 December 12, 2026, Atlanta, GA, USA 📝 Paper deadline: September 13, 2026 (AoE) 🔍 Review for us: if you work in formal methods, programming languages, automated reasoning, or ML for code, sign up at  https://lnkd.in/gcYv3baM 🎤 Speakers include Leonardo de Moura, Emily First, Vijay Ganesh, Patrick Li, Shan Lu, Baishakhi Ray, Bartley Richardson, Shubho Sengupta, Patrick Shafto, Dawn Song. 🔗 Details: https://lnkd.in/gTHuQrT2 #NeurIPS2026 #FormalVerification #AI4Code #NeurosymbolicAI

The Lean checker con-leche has a new offspring: con-ron. con-ron is con-leche in Rust, developed by Joachim Breitner. It closely ports con-leche to Rust, giving us a new Lean checker that relies on a different compiler and runtime stack. Joachim also used Aeneas to prove the core Rust implementation equivalent to con-leche. https://lnkd.in/g46X3Jb6

Ever wanted a pure-TeX implementation of the Lean kernel? No? Oh... Well, I made you one anyway. kernel.tex: https://lnkd.in/eUPTNWfX Please see https://lnkd.in/egxPqjXz for instructions for running it, and https://lnkd.in/eiqZdTj8 for its performance on the Lean Kernel Arena. It is nowhere close to checking all of Mathlib, but it actually does quite a bit better on the kernel arena if you give it longer time limits. Don't ask too closely about its provenance: I fed lean4lean into the maw of the shoggoths. I am looking forward to this announcement provoking more performant and/or obfuscated esolang implementations in the near future. #lean

I'm really pleased about the latest addition to the Lean FRO's computer algebra library Hex https://lnkd.in/gdrQCAPd written in #Lean: a full formalization of Brendan McKay's efficient graph isomorphism solver `nauty` https://lnkd.in/et_JNhp9. In some ways graph isomorphism is like SAT: in theory it's hard, but in practice you can do really well, and that is exactly what `nauty` does. `nauty` is one of my favourite algorithms. It's especially powerful when you realise that nearly all combinatorial objects can be encoded as graphs, in such a way that solving graph isomorphism solves the isomorphism problem for your combinatorial objects, too. Our implementation as the `HexGraphIso` library is a direct port of both the dense and sparse algorithms from `nauty`, along with complete Lean proofs that the port is correct (i.e. that it correctly determines whether graphs are isomorphic). It's certainly slower than the original C code, but appears to "just" be a small constant factor (~5x), and scales well on large graphs. We also provide (with correctness proofs!) functions to find generators for automorphism groups of graphs. You can read more about this in my blog post https://lnkd.in/e2Xn-X93 (but which only covers the dense algorithm), or in the Hex manual https://lnkd.in/eDXMuXcS (which also covers the sparse algorithm, which we finished today, and is much faster for most graphs).

I've loved chatting with Lance Fortnow about computational complexity over the years. He was reading "The Proof in the Code" when the Navier-Stokes news broke and has a nice writeup on his blog. "And now Lean, particularly in the Navier-Stokes papers, is being used as a time-stamp, a way to claim your theorem before having to write it up properly in an explainable way. Buckmaster even held back a result because it wasn't yet Lean verified. The way we even publish results is a-changing." https://lnkd.in/e48h7cmE