# The pressure gate

A new bounded maze mechanism lives at `/gate.html`. It complements the existing evolving maze at `/ecology.html`; both have direct links back to the nursery and the image trial. It does not replace or modify the frozen maze evolution engine.

The room contains two chambers, one connecting gate, a pressure plate and a movable crate. Agent or crate occupancy opens the gate. Leaving the plate removes pressure immediately. Reaching the goal requires pushing the crate onto the plate and walking around it to cross. Every attempted move, including a blocked move or wait, costs one physical action. Each episode has a 120-action limit.

The learner sees all geometry and current gate state. Occupancy is an authored input feature. It records open/closed frequencies separately for occupied and unoccupied readings, then uses those estimates in bounded breadth-first search over agent, crate and gate state. The planner and movement/push mechanics are supplied. A prior gate prediction is recorded before each move and compared with the actual state. This is a finite relational-learning puzzle, not an observation contract equivalent to the older locally sensed maze.

## Transfer comparison

Twelve independent seeds each supply one training layout followed by four generated evaluation layouts. Layout generation varies the gate row and goal and can reflect and transpose the room. Four matched controls use the same evaluation rooms:

- Retained: the trained occupancy table is frozen; evaluation cannot alter it.
- Fresh: learn a new table from the beginning of each room.
- Latched: an intentionally incorrect planning rule assumes an opened gate stays open. Its pre-action predictions use that same incorrect rule.
- No push: the crate cannot be moved, removing the required physical capability.

A separate author-side solver checks a true-rule witness route for each generated evaluation layout. Those routes never enter the learner. Physical training and evaluation actions, controller search transitions, and witness-search transitions are separate counters. The lease reserve is 2,040 physical actions per seed, a worst-case bound rather than claimed expenditure. Layouts within one seed are correlated.

The frozen protocol, source and independent evaluation seeds are in `experiments/frozen/conservatory-v7/manifest.json`. Development layouts 9101, 9103, 9109 and 9127 are excluded. A latched controller may happen to push the crate correctly, so it is not expected to fail every layout. There is no claim of open-ended mechanism generation, real-world tool use or AGI.

## Interaction

Play runs the controller; One step exposes one decision. Manual direction buttons pause automatic play and let you push the crate. Reset layout restores its initial knowledge, while Next layout carries observations forward. At completion, Play becomes Next + play. Reload reconstructs the saved journey from its initial table and actions; export includes that initial table, all actual observations, prior predictions and final knowledge.