FORGE / COMPACT EDGE MODELS
Review a small kernel pressure forecaster that learns only from completed evidence. Compare it with linear prediction and persistence before treating adaptation as an improvement.
The kernel model does not beat the linear reference in this experiment's mean observed-pressure error. Adaptation changes little; all results and missing windows remain available.
| Condition | Linear | Fixed kernel | Adaptive kernel |
|---|---|---|---|
| nominal | 0.02815 | 0.05196 | 0.05000 |
| demand step | 0.04040 | 0.06807 | 0.06788 |
| slow wear | 0.02992 | 0.05569 | 0.05461 |
| abrupt wear | 0.03982 | 0.07091 | 0.07062 |
| sensor bias | 0.05080 | 0.07907 | 0.07521 |
| quality gap | 0.02761 | 0.04911 | 0.04848 |
Values are mean per-run RMSE in a normalized synthetic pressure coordinate. They are not bar, a sensor accuracy specification or control performance.
Each report has a forecast-origin slider, the later observations and all four references. The sensor-bias case separates measured error from synthetic true-pressure error. Real data does not provide that truth oracle.
Four completed command/output pairs provide history. Eight declared command intervals describe the future recipe. Later commands cannot change earlier horizons, and incomplete target windows cannot update the model. The active dictionary is capped at 64 entries; this fit retained 15.
These are invented records and an experimental model. The offline tool has no control-write operation. It does not qualify a field process, prove sensor calibration or inherit the paper's stability guarantees.