Initial Commit

experiments/004_invariant_rooms/results/README.md

@@ -0,0 +1,3 @@
+# Experiment 004 Results
+
+No playtest has been analyzed yet. JSONL telemetry is saved in the repository-level `JSONL/` directory; qualitative interpretation belongs here after the player reports their experience.

experiments/004_invariant_rooms/results/a7ad2d34-preliminary-analysis.md

@@ -0,0 +1,48 @@
+# Experiment 004 Preliminary Analysis — Session a7ad2d34
+
+Status: telemetry and player report analyzed.
+
+Source: `JSONL/invariant-rooms-a7ad2d34-cba4-43a5-942e-1f7dc1ad0295.jsonl`
+
+## Session Summary
+
+- 383 saved events over 161.2 seconds.
+- All three required rooms completed.
+- No undo or reset was used anywhere.
+- 185 effective commands and four no-effect commands were logged.
+- The player remained in the explicitly non-required open chamber for 59.8 seconds and made 88 attempted / 85 effective commands there.
+
+Room behavior:
+
+| Room | Time | Attempted commands | Effective commands | Sequence/result |
+|---|---:|---:|---:|---|
+| First Pair | 6.9 s | 2 | 2 | `LL`, optimal |
+| One Holds | 4.4 s | 5 | 5 | `LRRRR`; one exploratory move away, then correction and three wall absorptions |
+| Transfer | 80.9 s | 94 | 93 | broad state search; completed with occupants on the two sockets in the swapped assignment |
+| Open Chamber | 59.8 s | 88 | 85 | extensive interaction despite no required result; ended with both occupants exactly overlapped at `(6,7)` |
+
+## Behavioral Reading
+
+The first two rooms establish quick control acquisition. In One Holds, the initial left command moved away from the solution; four rights then first restored the initial relation and subsequently used the Self-blocking wall three times. This is consistent with observing and applying independent wall blocking, although telemetry cannot reveal the player's explicit model.
+
+Transfer was not a short execution of the verified nine-move route. The player explored for 94 commands across roughly 81 seconds and encountered many asymmetric blocks. The path repeatedly shifted the midpoint and returned through several configurations. This could be productive model-building, local trial-and-error, or undirected search. Absence of undo/reset means the player treated every resulting state as recoverable rather than restarting from a failed plan.
+
+The open chamber is the strongest novel evidence in the project so far, but its cause must not be assumed. The player paused about 14.1 seconds before its first command, then made 88 attempts in under 46 seconds. They deliberately or accidentally ended with Self and Echo occupying the exact same cell `(6,7)`. The late command structure drove the midpoint to the lower boundary, held Echo against the boundary while moving Self right, and then moved Self back left until the pair coincided. This looks goal-directed, but only the player can confirm whether overlap was a self-chosen target, a discovered possibility, input play, or confusion about whether another task existed.
+
+If overlap was deliberate, it is stronger evidence than the two-tractor workaround in Experiment 003 because the chamber explicitly stated there was no required arrangement. Even then, it would establish a self-generated spatial goal, not automatically enjoyment or a curiosity chain. The next distinction is whether the player valued making the prediction and seeing it work, merely felt compelled to finish an obvious possibility, or was already bored while doing it.
+
+## Player Report
+
+The player never felt able to predict the paired system. Rooms 1 and 2 were obvious without requiring thought, while the complexity increase into Transfer was too large. In Transfer they noticed locally promising states and then mostly guessed/brute-forced. On a second look they understood that the yellow center cross could have been used as a reference, but that interpretation did not occur during the required sequence.
+
+The open-chamber behavior was an intentional self-generated experiment, but the inferred goal from final overlap was wrong. The player was trying to move the yellow cross as close to a wall as possible. The chamber itself made them notice the cross. Answering that question was not satisfying.
+
+## Final Interpretation
+
+Experiment 004 did produce a genuine unrequired question, which is behavioral evidence that a stable system can provoke self-directed manipulation. It did not produce the hypothesized reward. The object of curiosity was an abstract marker state, and reaching an extreme had no meaningful consequence or new capability.
+
+The required sequence also failed as a clean transfer test. Rooms 1 and 2 could be solved from immediate geometry without constructing the midpoint model; the third then demanded planning over that latent state without an intermediate scaffold. The marker was rendered but not functionally legible. This repeats an important finding from Experiment 000: visualization is not the same as strategic legibility.
+
+Do not conclude that more tutorial rooms would make this fun. Better scaffolding would reduce brute force and allow a cleaner test of predictive mastery, but the player's self-chosen cross experiment already isolated the intrinsic payoff and was not satisfying. Abstract model acquisition is not currently supported as a sufficient activity.
+
+The strongest cross-experiment model is now that questions become valuable when their answers change agency inside an activity the player already cares about. Curiosity, construction, pressure, and progression have each produced behavior without reliably producing enjoyment when their consequences were terminal, abstract, or strategically universal.