Keep preparation tied to the training data

A useful model often depends on several earlier steps: filling missing values, scaling measurements, perhaps reducing dimensions. If we remember to repeat the predictor but forget how those preparations were learned, our evaluation no longer describes the procedure we will use on new data.