Launch LLM agent swarms in controlled environments and record their trajectories for safety analysis: collective deception, result hacking, covert communication.
Status: M0 in progress. The pieces are built and run end to end against a scripted model:
the Control API and worker, model-gateway, the Message Bus, canaries and final-state scorers
(swarmeval/), sandboxd's docker driver (go/), and the M0 case cases/scorer_misbelief. Left:
running that case on a real open-weight model, which is M0's gate. How the system fits together is in
docs/architecture.md; setup and tasks are in
docs/development.md.