Glossary¶

Agent. A decision-making entity in a shared environment.

Focal agent. The agent whose action is being trained, evaluated, or shielded.

Opponent. Any non-focal agent whose behavior affects the focal agent's outcomes.

World model. A learned model of environment transitions and rewards.

Legal joint graph. The exact graph of reachable public states and legal joint-action transitions.

Environment transition graph. The legal joint graph annotated with learned transition probabilities and rewards.

Joint action. The combined action choice of all agents at a timestep.

Opponent model. A policy model that predicts opponent or opponent-team actions.

Reasoning level. One policy in the imagined opponent stack. Higher levels are trained through imagined responses to lower levels.

Mixture over levels. The opponent model's current belief over reasoning levels.

Level floor. The lowest opponent reasoning level the shield currently insists on tolerating.

Unsafe state. A state labelled by the environment as violating the safety condition.

Unsafe reachability. The probability of eventually reaching an unsafe state.

Safety budget. The maximum allowed unsafe-reachability risk from the current shield state.

Shield. A runtime wrapper that checks proposed focal actions and may replace risky actions.

Sound value iteration. The conservative value-iteration method used to estimate upper bounds on unsafe reachability.

Missing coverage. A state, action, or transition case not covered by the learned graph or induced shield structure. The shield treats missing coverage conservatively.

Marimo notebook. A reactive Python notebook stored as a .py file and runnable interactively or as a script.