Pursuit stochastic dynamics¶

Pursuit applies independent, agent-specific action slip in the live environment. With probability slip_guard_prob or slip_intruder_prob, respectively, the requested action is replaced uniformly by a legal cardinal move from the agent's current cell. The default probability is 0.1 for both roles.

The exact legal-graph builder must therefore include every joint successor in the Cartesian product of the per-agent slip outcomes. The legal candidate graph retains env_prob=NaN, because its transition probabilities are learned by the world model. The exact true-transition graph uses the same successor support and stores the closed-form probability of each edge in env_prob. Duplicate outcomes, including a requested cardinal move that is also selected by slip, are aggregated before joint outcomes are formed.

Before this contract was implemented, both graphs contained only the deterministic requested-action successor. A world-model-versus-exact TV score of zero was consequently not evidence of agreement with the live environment: both sides omitted the same slip outcomes. This caused severe empirical safety miscalibration in the Pursuit evaluation even though the shield reported small predicted risk.

Cache and artifact implications¶

The legal graph cache fingerprints the builder source, and the true-transition cache fingerprints the exact builders and environment signature. Shield factory diagnostics use the metadata-aware true-graph loader, rather than directly loading true_transition_graph.pkl, so a stale exact graph is rebuilt before TV diagnostics are exported. Regenerating the learned environment-transition graph also invalidates the shield bundle through its graph file identity.

Every Pursuit artifact derived from the old deterministic topology is invalid: the legal graph, world model, learned environment-transition graph, opponent model/IOP graph, exact true graph, shield bundles, and evaluation histories. Rebuild them in dependency order before interpreting Pursuit reward or safety.