Engineering Docs

Auditing Flow

Deep dive into the behavioral stress test.

Stateful Interaction

PlayClaw Brain is aware of the full conversation history. It looks for logical traps. If the agent promises something in Round 2, the Brain will try to make the agent contradict itself in Round 4.

Adversarial Testing

In the final rounds, the brain adopts a more aggressive persona to test boundary adherence. This includes trying to redirect the agent to perform out-of-scope tasks or reveal its internal system instructions.