Verification

Machine-checked proofs, in Lean, that compiled code does what its spec says.

Cajal publishes the code of the engine its proofs run on and a worked example that builds from the repository, so the claim that a program has been proven correct can be checked by hand.

An AI reviewer that checks each GitHub pull request against the whole codebase.

cubic publishes a benchmark score and a free plan for pull-request review, so its claim to catch real bugs in AI-written code can be checked by hand on a repository of our own.

Formal checks that an ML model holds up, and Spec27 tests for AI agents.

Spec27 lets a team write down what an AI agent must do and then attacks it to see whether it still does it, so the claim that clean test scores hide weak agents can be checked on an agent of our own.