Differentiable physics: long history and drift: research plan

Id	Question	Priority	Waits for	State
Part 6a	The recent-level filter on ensemble weather forecasts	Done	None	Experiment 111: the Kalman recent-level filter lowered the convex model's score by 0.151 points
Part 6b (N10)	Does training the Kalman filter jointly with the demand model, as one differentiable fit, lower the score at 1 to 2 days' lead beyond what the two-stage filter gives?	Done	None	Experiment 135: not supported. The untied joint fit scored 5.4 points worse, and the tied fit was within 0.013 points of the two-stage fit. No further joint-fit variants are queued
Shared day-level anomaly and sparse process noise (experiment 128)	Does one day-level anomaly state shared across the hour bands, with the process noise estimated sparsely from zero, fix the weights that the known-answer gate could not recover on real residuals in GAM-K's filter?	Done	None	Experiment 128: inconclusive, stopped at its known-answer gate. No rerun is queued, and no further per-substation learned-noise experiments after experiment 133 are queued

The later parts of the differentiable-physics series are planned one at a time, from the result of the part before.

Planned experiments

Experiment	Hypothesis	Relevance