What saved
work buys.
A cheaper forecast can keep Lumen planning for longer. Does that help both beds survive?
Three versions. Fifteen matched habitats. The same season and computation ceilings. An exposed development comparison, with the decision rule fixed before the run.
See where the allowance changes care ↓The whole season decides.
Loading frozen evidence…
“Viable together” is the fraction of 480 ticks when both beds remain inside their moisture ranges. Each version faces the same five conditions for each of three previously exposed seeds. These are development results; the fifteen habitats are not fifteen independent seeds.
Inspect every seed, season and work charge
Three seed means
All 45 complete seasons
Search work includes chosen and rejected evaluations. Model updates include the fitted model banks and any proposal validation. Logical requests equal actual transitions plus reused answers. Arithmetic, copying and memo bookkeeping are included in active wall time, but are not model transitions. No new runtime or energy comparison is claimed.
Watch the budget stop move.
Requests and work actually executed
Solid lines: executed transitions. Dashed lines: logical requests, including reused answers. The 12 million ceiling applies to the charged count: logical for the old allowance, executed for the new allowance and ensemble. The unchanged search and update reservations can stop planning below that ceiling.
These two first-seed seasons were selected before outcomes. All displayed motion and work curves come from recorded states. Seeking, replay and exports execute no physical actions or model forecasts.
Same rule. A different bill.
The optimized planner shares one waiting projection between both beds. Its two saved moisture answers expire at the end of each forecast. Here only the allowance reads a different counter; the loss, models, horizon and choice rule remain fixed.