CodexGameDev/experiments/004_invariant_rooms/hypothesis.md

Experiment 004 Hypothesis — Private Until After Play

Why this follows Experiment 003

Experiment 003 removed indirect automation but still reduced to obvious counter-loadouts. Its lattice rarely changed outcomes. The player nevertheless invented a two-tractor workaround while misunderstanding the heat system and still found the experience boring. This is important: deviation and problem solving are not equivalent to curiosity or fun.

Experiment 004 removes construction, upgrades, combat pressure, continuous steering, and numerical optimization. It isolates whether an unfamiliar but stable law can produce a valued cycle of observation, inference, prediction, and transfer.

Mechanic

One command affects two bodies:

• Self moves in the requested cardinal direction.
• Echo moves in the exact opposite direction.
• A wall blocks each body independently.
• Therefore ordinary moves preserve their midpoint, while an asymmetric collision changes it. Walls can absorb one half of a command and let the player ratchet the pair into configurations impossible in open space.

The interface does not state this law. It makes it legible through simultaneous tweened motion, trails, collision flashes, a connecting line, a midpoint marker, and a terse record of the last displacement. The player can undo or reset instantly.

Room Sequence

1. First Pair: an open horizontal arrangement. Two left commands solve it and expose opposed motion.
2. One Holds: one body begins against a wall while the other must move three cells. This isolates independent blocking.
3. Transfer: asymmetric internal walls require combining ordinary opposed moves with two different wall absorptions. The shortest unordered solution is UUUUURRDR (9 commands), verified by breadth-first search.
4. Open Chamber: no sockets and no completion condition. It exists only to observe whether the player has a question or prediction they want to test after the authored sequence.

Competing Interpretations

1. Inferring a stable unfamiliar law is intrinsically rewarding and produces a second question.
2. The reveal is momentarily interesting but becomes a short authored puzzle with no continuing possibility space.
3. The player enjoys spatial puzzle solving but not open experimentation.
4. The law is learned but applying it feels like laborious state bookkeeping.
5. The visualization is insufficient, so success comes from input search rather than a usable mental model.
6. The tasks are too easy or too short to expose the difference between satisfaction and compliance.

Evidence Priorities

Strong evidence for the target loop:

• a prediction stated or behaviorally tested before a required move;
• deliberate use of a wall after first observing asymmetric blocking;
• low-search transfer in the final room after exploratory earlier rooms;
• movement in the open chamber aimed at a self-chosen configuration;
• a concrete question about a consequence not required by sockets;
• wanting another situation because of the law rather than merely another puzzle.

Ambiguous evidence:

• completion;
• resets or many moves;
• finding an unintended route;
• spending time in a room;
• any claim that the mechanic would or would not be fun with more content.

Negative evidence:

• immediate mechanical input search without model formation;
• stopping as soon as sockets are filled with no desire to predict anything else;
• describing the law as understood but exhausted;
• finding state planning tedious even after the behavior is clear.

Instrumentation

Log every command with before/after coordinates, requested direction, each body's displacement and blocked state, midpoint before/after, move count, undo depth, and completion state. Also log resets, undos, room entry/completion, time and move totals, open-chamber moves, visibility changes, and saves.

Telemetry cannot distinguish thoughtful prediction from trial-and-error by itself. Pair it with the player's report and inspect pauses, undo patterns, and whether wall interactions become more intentional over time.

Questions After Play

Ask only after receiving the JSONL:

1. At what point, if any, did you feel you understood what the paired movement and walls would do before pressing a key?
2. In the last required room, were you executing a plan, testing local guesses, or searching inputs until something worked?
3. Once the required sequence ended, did you have any result you wanted to produce or any question you wanted answered?

Do not ask whether “discovering rules” is fun; that wording would encourage agreement with the hypothesis.