If the user provides us with the input feature descriptions, then in the Predictor we will be able to verify that the data set for prediction contains all the necessary columns at design-time (not currently implemented but should be easy to do).
We need this information to figure out the output structure of the Predictor, e.