It seems like a question of: how expensive is a system that’s corrigible on some universal core for reasoning? If the system is very expensive (e.g. because we have to think for a hundred hours for each data point), then you may not be able to get enough data to learn it.

The real bad case is when you have to think about the entire hypothetical subtree beneath a question, in order to find a corrigible decomposition, because (by assumption) that subtree is too large for you to consider explicitly.