Evidence at a glance #
Based on the linked sources. The author's reported result and independent confirmation are separate checks.
- AI involvement
- Reported models: Claude. Tools: Lean 4, nanoda, Comparator, in-toto, SLSA. The writeup below describes the work attributed to AI.
- Reported result
- Research. This is the directory's source-based outcome classification.
- Independent confirmation
- Not reported.
- Human work and limitations
- The repository explicitly states that it is not a mathematical review and does not verify the correctness of the underlying proofs against the original papers, only that the Lean code compiles and checks.
Original sources
What was done #
This repository provides a reproducible QA audit of OpenAI's Lean 4 proofs for Navier–Stokes and Euler equations. It rebuilds artifacts from source, validates them with independent checkers, and tests the pipeline against mutations, while documenting all findings and limitations.
The Scientific QA Evidence repository provides an independent audit of OpenAI's Lean 4 formalizations regarding Navier–Stokes and Euler equations. The author, Cristian Guerra, applies software QA methods to rebuild published artifacts from source, recheck them with independent tools, and verify the pipeline's integrity. The project includes a mutation-testing campaign on nine numerical models, which flagged 54 deliberately broken variants while missing three tiny changes. The author documents that the repository does not claim new mathematical results or a Millennium Prize, but rather provides a public record of commands, versions, and hashes. The audit process identified and corrected several internal mistakes, such as a results verifier that initially accepted logs containing Lean sorry warnings and failing builds. The project also includes a Lean proof attempting to connect OpenAI's theorems to the Clay Institute's official problem wording, though the author notes this transcription remains subject to expert review. The work was co-authored with Claude Opus 5.5.
Reproduce it #
Follow the steps in the repository's documentation, specifically the FROM_ZERO and B5 build instructions, to rebuild the proofs from source using the provided hashes and commands.