Blog
Theorem research blog
theorem.devCrawled Oct 14, 2026
Theorem research blog
Fractional proof decomposition fuses partial evaluation and property-based testing to scale testing compute logarithmically with bug rarity, instead of linearly.
lf-lean is a verified translation from Rocq to Lean of all 1,276 statements in Logical Foundations, done by frontier AI 350× faster than humans.
Theorem builds products to make software correct, understandable, and secure.
Formal verification and AI — research and engineering from Theorem.
A new startup called &Truth is joining the race to automate formal verification, a painstaking process that uses math to prove software works as intended.
Anthropic’s Frontier Red Team uses Claude to perform formal verification on **123 open-source libraries**, proving security properties at the assembly (RISC-V) level. The AI autonomously generated proofs for **239 entry points**, uncovering long-undetected bugs despite trillions of prior fuzz-test runs. These proofs guarantee no crashes on inputs under 64 MB for single-threaded static binaries under Linux, covering memory-safety violations and assertion failures. The team also published proof artifacts and a Claude skill via collaborator **Theorem** to accelerate adoption. While not exhaustive, this marks the first large-scale AI-driven formal verification of production software.
Theorem presents lf-lean, a verified translation from Rocq to Lean of the statements in Logical Foundations. Theorem's February 2026 research describes task-level specification generation and verification at scale, positioning the work as company research into software that can be checked formally.
Theorem explains fractional proof decomposition, a method that breaks end-to-end software properties into smaller, testable properties. In an October 2025 research post, the company applies the method to an approximate top-K sampling bug and says it trains models to reason about program correctness.