Claude proved 239 library entry points crash-free, using about 180 million lines of Rocq

What was actually proven

Anthropic's post argues that LLMs make formal verification feasible at scale. The proofs are in Rocq, over assembly for RISC-V binaries, so they do not trust a compiler. The spec for each library comes from its fuzz tests: the program never crashes for any input under 64 MB. Assumptions include a single-threaded static binary, fixed memory and Linux.

Results: 239 entry points across 123 libraries, about 180 million lines of Rocq in total, hundreds of bugs found, sometimes fixed by Claude before proving. A Rust Python interpreter matched CPython on 99.95% of 1.7 million random programs. The authors list limits: the trusted base includes the Rocq checker and the hardware and OS models, and an agent could make a theorem provable by weakening the program, so patches were audited.

Scepticism about what the proof is against

One commenter quoted the appendix on scaling:

"The proofs written by agents are just as 'good' as any other proof of the same theorem, in that they prove the equivalence theorem. However, it remains to be seen whether agents can scale this verbose proof style directly to more complex programs." — a commenter quoting the paper's appendix

Another asked whether fixing crashes found by a fuzzer and then proving the fix is pointless. A third asked how anyone knows the statement models the real program. If you maintain a C library, the takeaway is narrower than the headline: a crash-freedom proof is a bar fuzzing can approach but not certify.