A baseline trainer for RF and pedestal response data
The RF and pedestal data interfaces now connect to a reusable ensemble trainer. It accepts the declared features and outputs of a reference dataset, so the numerical training pipeline can be tested before choosing new physics closures.
Keep the holdout intact
Failed labels are filtered inside each saved split. They do not trigger a fresh split, and related cases stay together. For profile labels, cells become training rows only after cases have been selected. The caller must explicitly choose the layout that combines case features with the supplied radial coordinate. There is no interpolation or coordinate conversion.
Normalisation is fitted on the training rows alone. Each ensemble member keeps its best validation checkpoint, and the test set is scored after fitting. Errors are reported in each output’s declared units, alongside a constant training-mean baseline. Ensemble spread is a diagnostic here; it has not been calibrated as a coverage guarantee.
Inference carries what it needs
The saved model includes its normalisation, feature schema and training-input box. Reloading checks both the model and metadata hashes and does not open reference datasets or solver files. The box is the range of the training points, not a certified domain. Loaded models can be passed to the existing response guard with an explicit fallback and physics bounds.
Synthetic tests check the pipeline: 26 synthetic baseline-pipeline tests pass in full and Tier-A-only modesC-145. They include gradients, saved splits, unit-preserving metrics, model tampering and deterministic replay. Case and profile fits improve on the constant baseline in these numerical fixtures. That is an optimiser check, not evidence of RF deposition or pedestal prediction accuracy.
The trainer is an unconstrained baseline. No production RF/pedestal dataset or physics model has been trained, and no positivity head, power partition or pedestal criterion has been chosen. Selecting reference solvers or data is still the next modelling step; then the baseline can guide accuracy targets and the physics structure needed for the final model.
Technical details → Response training.
Technical details → Surrogates › response training