Algorithmic risk assessment in criminal justice is defended on a comparison that is easy to state and hard to refuse: the alternative is not neutral judgement but the judgement of magistrates, which is inconsistent between individuals, sensitive to the time of day and demonstrably shaped by the same social patterns the algorithm is accused of encoding. Where the comparison misleads is in treating consistency as the principal virtue at stake. A model applies a rule derived from historical outcomes to an individual, and its errors, unlike a magistrate's, are correlated: every person the model misjudges in a particular way is misjudged identically, and there is no second opinion anywhere in the system. Distributed inconsistency and concentrated systematic error are different failures with different remedies, and the case for the algorithm establishes only that the first is bad.