Logistics model diagnosis
Lumen Dispatch asks whether a delivery-time model deserves a limited pilot. Compare Ridge with a constant baseline on the same held-out records. Keep error units explicit, inspect signed residuals and finish with a checked binary recommendation. The fictional fixture supports a sampling-based comparison, not a claim about deployed performance.
A parameter search compares a declared set of model settings using the same validation plan. It doesn't turn the test set into another tuning table. You will fit a small search, read losses in their natural units and evaluate the selected model once on a held-out partition. Pipeline parameter names include the step name and two _ characters. The scorer also matters: maximizing a negative loss is equivalent to minimizing its positive value. If candidates tie, use a rule chosen before examining the test score. The smaller model wins the tie in this lesson because the operational requirement says so.
train_mae = mean_absolute_error(reg_y.iloc[tr], delivery.predict(X.iloc[tr]))
Training error is a fit diagnostic, not an estimate of unseen performance.