A promise
of later.
Lumen predicts that it will close a valve soon. Then it chooses another reading, and “soon” moves forward. The forecast sees a rescue that the figure keeps postponing.
A recorded trace of four previously exposed continuations. The candidate repaired one failure but made full seasons worse. Ensemble care remains the default.
Follow the plans ↓When does “later” arrive?
Loading recorded plans…
Horizontal: when the plan was made. Vertical: when its selected-model forecast expects the next operation. The diagonal means “act now.” Gaps mean no matching operation was predicted within that plan’s 48-tick horizon. Dashed horizontal lines mark recorded operations within the replay window.
Model weights are heuristic support, not calibrated probabilities. Every row shows the same chosen initial journey; the original forecasts allow each model to invent its own later actions.
A local repair can still be a poor policy.
The candidate makes all three models evaluate the selected model’s same follow-up script. The initial committed journey is identical. Afterward, each controller uses its own new local observations and decisions.
These four cases come from the two fixed v0.14 examples: the original choice and the previously lowest-cost initial journey in each. They are explanatory cases, not new validation. Playback performs no simulation.
The candidate did not earn a fresh test.
Before the pilot, we required a nonnegative mean care difference to proceed. The candidate failed on all three exposed seeds. It was retired without changing its settings or consuming fresh evaluation cases.
Let a future reading change a future plan.
The next candidate should share actions while observations remain indistinguishable, then allow different follow-ups when an actual reading separates the possibilities. That is a hypothesis to test; neither the diagnosis nor this rejected fixed-script rule proves it will improve care.