How far a
promise reaches.
A forecast can make a poor journey look good. Is the problem what happens along the way, or the water it promises to save?
Sixty fixed, previously exposed checkpoints. Twelve possible journeys each. Removing the reserve penalty changed no choices. Looking another 48 ticks ahead changed 35, without improving care.
Take a score apart ↓What makes it seem worthwhile?
Loading frozen evidence…
Lower scores are preferred within a rule. The first 48 predicted ticks are identical. The reserve estimates water shortage over the rest of the season; the extension simulates only 48 extra ticks. These are different terminal assumptions, not equivalent horizons.
Inspect the water estimate in each fitted model
Current weather and each fitted model are held constant, with both shades deployed. Need and available water are estimates, not a feasibility proof. Support weights are heuristic, not calibrated probabilities.
Actual 96-tick costs and viability in the table come from committing the initial journey, then resuming unchanged ensemble care. They were unavailable when the forecast chose. Viability is the fraction of ticks when both beds are inside their moisture ranges.
The same beginning can lead elsewhere.
Both panels show recorded physical outcomes. Left: an offline observer executes the selected fitted model’s entire predicted action script. Right: Lumen commits only the same initial journey and then replans from local readings. The script is a diagnostic intervention, not a new live ability. Playback performs no simulation.
No scoring change earned a pilot.
Compare seed means and the frozen nomination rule
To earn a separately frozen full-season pilot, a rule had to lower mean 96-tick regret, raise mean 96-tick viability, and avoid a viability loss in at least four of six seed means. Neither qualifies. No fresh evaluation followed.
Regret is realized cost above the best of the twelve recorded continuations at each checkpoint. These are retrospective 96-tick opportunities, not full-season policy results or fresh generalization estimates.
There is work we may be able to reuse.
Two complete searches from the earlier two-decision candidate. Every model call still ran. Exact fitted-state/action matches measure reuse opportunities; they do not measure an achieved speedup. Key creation, storage and lookup have costs. Each instrumented journey matched the original search result exactly.