Agentic Proving for Program Verification

In three linesClaude Code evaluated on CLEVER (Lean 4 benchmark) generates valid specifications for 98.8% of problems, certifies 87.5% of implementations, and achieves 98.1% success on end-to-end program generation and verification. Study reveals mismatch between current benchmark difficulty and modern agentic prover capabilities.

Summary generated by Claude — human-verified