I wrote down my ideas about how to test whether a post-hoc explanation (from something like LIME, Shapley values etc.) is accurate.

Download