Where a
plan changes.
Two beds can share one projection of the future. But when that future arrives, Lumen may choose a different journey.
An exact computation improvement and four fixed explanations. Search runs faster; the question of better care remains open.
Inspect the first fork ↓Wait once. Read both endpoints.
The old forecast simulated waiting separately for each bed. The new version follows the same waiting sequence once, saving each bed’s predicted moisture when its travel-and-delay time arrives. The saved answers expire with that forecast.
Loading frozen evidence…
Inspect all ten timed pairs and memory samples
Warmups and comparison runs are charged in the evidence. Timing covers each complete twelve-root search; output comparison and explicit garbage collection occur outside its timer. Heap changes are before/after samples, not peak memory. The memo retains only two moisture values and one working model state; it has no persistent cache.
The original logical work allowance is still charged, so budget stops remain identical. This is an optimization of the unpromoted two-decision planner. It does not establish better care, lower energy use or a speedup in other workloads.
What changed the next journey?
Inspect local readings, fitted models and candidate scores
Readings since the common checkpoint
Models used to choose
The same ordinary journey menu
The shadow shares the initial reading, weather, equipment and actual actions. Later moisture and tank readings are withheld throughout the intervening history. Its proposed journey is then tested using the real memory and real readings. Support weights are heuristic, not probabilities.
Both panels show recorded physical outcomes from the same decision checkpoint to the same end tick. The intervention commits a short action sequence, then resumes ordinary ensemble care. Real readings continue throughout. Playback performs no simulation.
“Commit the current choice” checks whether commitment itself changes the result. Compare it with “Choice without later readings” to isolate the proposed journey. Script commitments stop at the next valve, shade or drain operation, capped at sixteen ticks.
A local repair is not a general rule.
These four cases were chosen from earlier exposed evidence before this run. The successful alternatives are hindsight selections, unavailable to the learner. Percentages compare the common remaining window within each case; they are not complete-season policy results.