is there any interpretability result you would actually bet your product on

not asking what is interesting, asking what is reliable enough to act on in a real system. genuinely curious if we are there for anything yet.