Public repositories provide machine-checkable Rocq proofs and verifier tooling, but cover less than the publisher-reported internal run.
Anthropic and Theorem publish machine-checked proofs for C libraries
2026-10-22Verified announcementTRIAL
Anthropic and Theorem released MIT-licensed Rocq proofs and verifier components for selected C library fuzz entry points. Anthropic reports a larger internal run, while the public proof repository currently covers 60 entry points across 25 libraries. The guarantee is bounded to modeled static, single-threaded RISC-V Linux processes and each fuzz target's stated properties.
What this means for you
For teams evaluating AI-generated verification, the public artifacts allow proof checking rather than trusting a benchmark score. The gap between the reported internal scale and public subset, plus the narrow execution model, means this is a reproducible pilot rather than a blanket security guarantee.
Evidence & related material
Technical details & scores
| Public Artifacts | 60 fuzz entry points across 25 libraries |
|---|---|
| Publisher Reported Internal Run | 239 entry points across 123 libraries; about 180M lines of Rocq |
| Proof System | Rocq with modeled Linux process semantics |
| Scope | static single-thread RISC-V processes, musl and modeled Linux behavior |
| License | MIT for the cited public repositories |
| Time Basis | official article and repositories give 2026-10-22 date precision; strict hourly placement is unavailable |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 2
Verification cost 3
Risk 2