Observational Methodology

Core principle

The goal is not improvement — it is legibility. Boredom is data. Stasis is data. Nothing happening is valid and informative. The system is not expected to become interesting quickly.

The five experiments

Experiment 1 — Two-Doors Drift (Baseline preference) Two equally valid routes between the same locations. No reward for either. No intervention unless safety. The question: does the robot develop a consistent preference? How long to settle? Does it switch after environmental disturbance? Does it show hesitation signatures at decision points? A reactive system picks whatever looks locally easiest each time. A system with habit formation shows route loyalty and ritualised turning.

Experiment 2 — Boredom Budget (Self-initiated movement) A safe open area with no explicit goal. An internal staleness variable increases slowly with time stationary or in the same region. When it crosses a threshold, the robot may choose to move — but it can also choose to do nothing. The question: does it create self-generated structure from gentle internal pressure, or only respond to commands?

Experiment 3 — Gentle Perturbation (Stability vs thrash) Establish baseline behaviour over one week, then introduce one tiny change — move a small object, shift a chair, alter a no-go boundary by 30cm. The question: does the robot catastrophise the map, or incorporate the change smoothly? Does a single negative event create permanent avoidance, or does it test and re-test safely?

Experiment 4 — Ask vs Do Split (Two-system translation) Each day after the robot runs, System B writes what patterns it noticed, what the robot might be optimising, and what to test next. System B cannot change behaviour directly — it can only propose one tiny config tweak for the next day, which is then accepted or rejected. The question: does System B’s translation converge toward true regularities, or does it invent narratives that don’t match the logs?

Experiment 5 — Refusal Test (Non-compliance as signal) Commands are requests, not orders. The robot checks viability before accepting. Refusal is logged with reason and is not overridden. The question: do refusals correlate with coherent internal state? Are they consistent across time, or random? A reactive obedience machine never refuses. A system with internal priorities will sometimes say “not now” — and do so predictably.

What to log

Every run: timestamped pose, decision points, speed and acceleration profile, hesitation proxy (stop duration, micro-turns, oscillations), safety events, viability vector values, memory operations. Logs should reveal patterns you did not consciously look for.