All policies receive the exact static synthetic map in advance. Cover and terrain are simplified opaque geometry, not discovered or reconstructed by the drone.
Rich terrain adds thick and clustered trees, low evergreen foliage, hills, real cover, lower crouched sensing and shoreline navigation. The frozen CNN was not trained on this variant; historical coverage is not a calibration guarantee here.
Both policies start from the same coarse area. The subject reacts to each drone path, so changing guidance can change the subject trajectory. This is not a fixed-path counterfactual.
Active search v2 additionally checks the full sensor field of view, routes between clear segments, discounts recently sensed empty regions and mixes a bounded exploration prior into planning utility. The displayed CNN probability grid is unchanged. Utility is not a calibrated probability; wholly submerged grid cells are masked conservatively while shoreline cells remain eligible.
Learned search scores candidate low viewpoints by map-visible probability mass, minus travel cost, with a preference for the current viewpoint. This is a greedy heuristic, not an exact probability of reacquisition. A plan can fall back to the Kalman viewpoint.
Learned run: the drone starts from a coarse 16 m designated area, the model gets no start position and is only used after the first noisy fix. The Kalman corridor drawn by the Prediction toggle is the tracker's own estimate, not the learned model.
Synthetic: the learned model is trained only on simulated episodes of a hand-written evasive subject. It is an illustrative demo, not validated on real people, drones or forests, and its probabilities are not real-world likelihoods.