How do we / should we handle columns that could give additional information during training that are not present for prediction? Example would be to use biological quantities that are hard / expensive to measure in addition to normal patient information to predict disease outcome. One could then build a pipeline that uses a LearnerCV to predict the expensive biological quantities first, and integrates that into the prediction of outcomes. In resampling, the expensive quantities would then need to be handled like target columns to some degree, insofar as they should be absent during prediction. Does something like this exist already?