I'm starting to believe that maybe it is better not to talk about data at all, perhaps it is best to ask users to directly return a cost (or a vector of costs if it makes sense for some algorithms) in this way data could in principle change during inference in a user defined way, what do you think?
Originally posted by @francescoalemanno in #5

This sounds like a good idea! I have to think about this and look also at the methods that would be nice to be implemented in the future to see if this is flexible enough...