Something has to
hold the door.
A room has several ways through. A distant plate can hold a gate open, and familiar paint can conceal a different release delay. Carry observations into the next room, then check whether they still apply.
Move into the crate to push it. Manual moves pause automatic play and still supply observations. A blocked move costs one action.
Does the rule travel to another room?
Twelve seeds. Train on individual gates, then test independent changes to paint and wiring in larger rooms with two or three alternative passages. Each controller has 180 physical actions and 30,000 search transitions.
A curriculum that changes the room.
Mutate an archived room, evaluate its learner and up to two distinct donors, and keep solvable pairs. Three seeds, six generations each, up to three gates. These are selected training cases, separate from the comparison above.