What exists
- A threat model that says what the system promises against each kind of attacker, and what it does not.
- A suite of adversarial scenarios kept as permanent regression tests.
- Versioned decisions and an experiment register.
What is being tested
Whether a workflow that binds evidence versions, policy, date and jurisdiction to each decision makes fewer unsupported claims than a strong ordinary baseline, at an acceptable cost and review load.
Why it is in the lab
It is the most general form of the lab’s question: when is a machine allowed to say something, and when is it allowed to act.