AI companies are building tools that use mathematical proofs to check software, offering a stricter alternative to tests that can find bugs but cannot show they have found them all. A new entrant, &Truth, is building a verification system on Lean, a proof assistant that checks whether a proof follows its rules.
Watch Desk analysis
What happened
A Latent Space article examines formal verification tools from &Truth, Harmonic, Theorem and Axiom Math. It describes a workflow in which AI helps construct proofs and a separate proof checker checks each step. The goal is to verify that software meets specified requirements, rather than treating an agent’s tests or self-review as a guarantee.
&Truth says its first tool, &Atlas, maps open-source C code. Co-founder Clark Barrett told Latent Space that one intended use is finding code that is not memory-safe and identifying problems to fix. The article also describes cvc5, a reasoning engine co-created by Barrett, as being used for specification analysis and test generation in Amazon’s Kiro development environment.
Why it matters
Tests can expose failures, but passing them does not establish that every relevant failure has been caught. A machine-checked proof can offer a stronger kind of assurance, provided the requirement being proved is precise and the system really checks the result. That distinction matters as AI agents produce more code for people to review.
The approach is not simply “ask AI whether the code looks right”. In the examples described, AI helps produce formal proofs, while software such as Lean checks their logic. That could make rigorous verification more practical, though the article does not establish that these tools can yet verify arbitrary software at scale.
Our read
This is a more compelling answer to AI-generated code than asking another model to give it a reassuring once-over. But proofs are only as useful as the requirements they prove, and a polished mathematical certificate cannot rescue a badly specified job. The interesting shift is towards making verification part of the coding workflow, not a ceremonial final glance.
What to watch
- Whether &Truth publishes working details or results for &Atlas.
- How much code these systems can verify, and at what cost.
- Whether developers can specify requirements precisely enough for proofs to be useful in everyday projects.
Discussion spark: Should AI-written code need machine-checked proofs before it is trusted in important software, or are strong testing and human review the more practical standard?
Sources and evidence
- The War Against Slop and an Emerging Weapon to Fight it: Formal Verification (20 November 2026, 21:18 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.