Formal verification startups pitch proofs to check AI-generated code
Original title: The War Against Slop and an Emerging Weapon to Fight it: Formal Verification
&Truth, Harmonic, Theorem and Axiom Math use Lean and solvers so proofs, not tests, decide when code is done.
A Latent Space interview with &Truth co-founder Clark Barrett surveys companies selling formal verification as a check on low-quality AI-generated code. Barrett contrasts tests and model judges, which cannot show every issue was found, with Lean proof checking. &Truth’s Atlas starts with open-source C and aims to flag memory-safety problems; Harmonic, Theorem and Axiom Math also build Lean-based provers.
Unconfirmed: Claims that Lean proofs settled Smoktunowicz’s Gap Theorem, Fermat’s Last Theorem, or the Riemann hypothesis are attributed to Vals or Levent Alpöge, not independently confirmed here.
- Barrett says cvc5 is called billions of times a day in Amazon production pipelines, including Kiro.
- Axiom Math uses proof attempts as correctness feedback when training AxiomProver.
- Harmonic’s Aristotle has AI write proofs and Lean check each step.
- On 36 small tasks, Anthropic said Opus 5.5 used about 40% fewer tokens than Opus 5 for a 50% solve rate.
- AI models
- GPT-6.2 AstraClaudeOpus 5.5
From the original article
6 min read at latent.space
Proving software correct with math has traditionally required specialist work. Now AI and Lean are helping automate the process.
By Richard MacManus · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.