Notes on proving AI-generated code safe.
The engineering log, in public-ready form. How the verifier works, what our studies actually show, and where a careful reader should push. Written for the person who will try to break it.
Why we verify AI-generated code instead of scanning it
A scanner asks whether code matches something already known to be bad. A verifier asks the AI to declare what its code does, then proves those declarations hold. One approach chases an unbounded list. The other closes it.
1,572 cases, zero misses
The method, benchmarked. Every vulnerable case in the OWASP Benchmark's six in-scope categories blocked at full coverage, with the false-positive floor decomposed instead of hidden, and the instrument published.
Recent writing
F1 = 1.0, and 62.5 points to zero
Two numbers do most of the work when we describe Acutis. Here is exactly how we got each one, exactly what it means, and exactly where a skeptic should push.
Zero enumeration: a verifier with no rule list
No function lists, no patterns, no stdlib defaults. The verifier is name-blind, so a framework it has never seen works on day one. Why that is a moat, not a gap.
Proves safe, not catches bugs
A gate you want to pass, not an alarm you learn to ignore. What changes when security verification runs inside the AI's loop, before code reaches a person.
A proof bundle, not a risk score
Each verdict emits a machine-readable record: the property path, the trust assumptions graded by strength, and nothing assumed in silence. The evidence an auditor can read.