Adversarial Audit

If a human keeps finding your bugs one at a time, you don't have a system with checks and balances — you have a person doing QA for a robot. That doesn't scale, and it burns their trust.

The fix is cheap: point an agent that built none of it at the thing, hand it the failure history, and tell it to be skeptical rather than reassuring.

Ours found ten bugs, four critical, in about six minutes — after days of the human finding them one at a time.

Why a fresh agent beats the one that built it

An agent reviewing its own work grades its own homework. It knows what it meant to do, so it sees what it meant to do. It is also, subtly, invested in the work being good.

A separate agent — separate context, no memory of writing the code, no stake in it being correct — has none of that loyalty. It is the only party in the system with no interest in the answer being "it's fine."

How to run one

1. Give it the failure history — this is the highest-leverage input

Don't just say "review this code." Say: