A reading
forks the plan.
Travel together. Read locally. Then consider different next journeys only when the readings give a reason to separate them.
A bounded two-decision experiment. Every branch below is a forecast from fitted models. The recorded continuations show what happened when Lumen replanned after real readings.
Inspect a fork ↓What would change the next choice?
Loading frozen evidence…
Moisture is grouped into two-percentage-point bands; tank level into 0.05-unit bands. The entire sequence of local readings along the first journey defines a branch. Models sharing a branch must share all forecast actions. Support weights are heuristic, not calibrated probabilities. Only the first journey is committed; later real readings trigger a new search.
Compare the recorded continuations.
These are recorded observer states. Playback performs no simulation. The fixed short cases and the first seed’s steady and compound seasons were retained regardless of outcome; they are development examples, not fresh validation.
The whole season decides.
Before running, we required a complete pilot with a nonnegative mean paired viability difference to earn a fresh comparison. Every season counts, including those that reach the computation limit. Candidate settings and observation bands are frozen.