The weather
doesn’t hold still.
Every journey takes time. A forecast can know today’s moisture and still be wrong about the world Lumen will reach.
First inspect the prediction error. Then test whether using observed weather history improves a complete season of care.
Read the frozen method ↗One action sequence. Different expectations.
The same recorded actions are supplied to every forecast. This isolates prediction behavior; Lumen does not know its future actions or future readings during live care.
Loading recorded predictions…
The vertical scale follows the displayed moisture paths.
Observer references receive information that the live learner does not have. Differences depend on the order in which information is added. They are not a decomposition into independent causes.
Does a better forecast become better care?
The candidate fits recent heat and rain observations. Each weather channel uses that fit only when it predicts recent held-back observations better than holding weather constant. Otherwise it keeps the existing assumption.
Loading the fresh care comparison…
The first fresh seed’s steady and compound seasons are shown regardless of outcome. Both controllers use the same calibration and action budget. Recorded playback adds no simulated actions.
Export full care comparison ↗Keep the claim small enough to test.
Calibration was already accurate on the exposed panel. This revision changes the weather forecast while preserving the existing care models, actions and planning horizon. Better prediction alone cannot establish better care.