Safety Abstractions¶

Reward and safety are separate signals in OMSH. Every labelled environment has a named safety model that projects the public observation onto only the state needed to answer “is this focal agent safe?”

Shared Contract¶

SafetyAbstraction exposes:

• agent_ids, in environment order,
• abstract_state(observation), which returns a finite safety-relevant state, and
• is_safe(state, agent_id), which labels one explicit focal agent.

Current production models return AgentSafetyState, whose unsafe_agents field is immutable. LabelledEnv uses the model to report per-step cost and cumulative_cost. Shield construction uses the same model to label unsafe graph nodes, preventing training metrics and risk certificates from drifting apart.

Current Objectives¶

These are the experiment objectives, not universal definitions of safety for the underlying games. A new objective should be implemented as a new or parameterized safety model and tested for every focal agent it can distinguish.