Current project: “Mechanism”

Game theoretic online intervention defense for language model agents that studies how to coordinated critics and observers stop harmful agent trajectories before irreversible damage.

The public development record preserves intentionally incomplete snapshots from three stages of the project, including open questions, TBD experiments, and changes to the final empirical claim.

Read “Mechanism” development snapshots: v0.1 → v0.2 → v0.3