Well, you're right about that distinction-- but from your comment "only on Tuesdays" I understood you were talking about behaviour modification some time after "load time". Also keep in mind the first paper cited in the article (profile-guided inference) actually points out that in most cases, "load time" can be inferred nearly statically-- as in, you would not need to run a test suite, you would only need to "load" the base code. This is a feasible task, in most cases.