A growing group of AI companies is applying formal verification—mathematical proofs of software or hardware behavior—to code written with AI. Unlike tests, which can reveal known failure cases without proving that none remain, a proof assistant checks every logical step against a stated specification.
&Truth is building a verification system on Lean, an open-source language and proof checker, while Harmonic and Axiom Math also use Lean in automated theorem-proving products. Theorem has worked with Anthropic on scaling software verification with language models. On 36 small verification tasks, Anthropic reported that Opus 5.5 reached a 50% solve rate with about 40% fewer tokens than Opus 5.
The approach still depends on humans defining the right requirements, and proof generation can be expensive. But the developer workflow could shift from reviewing every generated line toward writing specifications and checking machine-verified outcomes. Security is an early target: &Truth says it wants to inspect C code for memory-safety problems, while its cvc5 reasoning engine is already used in Amazon’s Kiro development environment.