Probabilistic Shielding¶

Probabilistic shielding is the safety layer that sits between a learner and the environment.

The learner proposes an action. The shield estimates whether that action respects the current risk budget. If it does not, the shield replaces the action with a safer pure action.

What Risk Means¶

Risk is the probability of eventually reaching an unsafe state, not just the probability of being unsafe on the next transition.

The shield computes conservative upper bounds on unsafe reachability with sound value iteration. These bounds become certificates used at runtime.

Runtime Behavior¶

The canonical runtime wrapper is TransitionShield.

At each step, the shield:

1. reads the proposed focal action,
2. checks the relevant opponent-model levels,
3. estimates the risk of available focal actions,
4. keeps the proposed action if it fits the budget,
5. otherwise replaces it with a robust safer action,
6. updates the remaining budget after the realized transition.

The base environment action space is unchanged. The shield changes what action reaches the environment, not what the learner is allowed to propose.

Learned Shield and True Shield¶

The learned shield uses the learned environment transition graph and opponent model artifacts.

The true-policy exact shield is a baseline that uses exact transition structure where available. It is useful for comparison, but it is not the default learned pipeline.

Telemetry¶

The shield can expose debug information such as:

• whether an action was overridden,
• proposed and final action risk,
• current level floor,
• missing coverage,
• budget before and after the step.

Experiment notebooks use these fields to summarize shield behavior alongside reward and safety metrics.