Who audits the auditors?

Make trust answer to evidence.

We are an independent group of Stanford students,
led by Howie Xu, researching AI-agent security.

The counterexample. A claim, an interruption, a revised path.

When evaluators have incentives to cheat,
what evidence makes their findings trustworthy?

Read the paper

Working paper 01 · Formal analysis and synthetic experiments.
Summary & replication files

Open the evaluation workspace →
Private pilot · Run, inspect, retest.