The Sentencing Algorithm

The problem: a county buys a risk-scoring model to guide sentencing — faster, cheaper, statistically tuned; file the case with the punishment and machine tools together.

Open the file. Undisputed: the model exists, the inputs are listed, the county signed. Disputed, flagged: every accuracy claim — those are fact questions, sent to evidence with the flag showing. The affected: defendants scored, victims, judges lending the model their authority, the neighborhoods the proxies describe. Now the arguments at full strength. For: the deterrence family's case — sentences made consistent, incapacitation aimed by evidence rather than mood, and a judge's bad morning no longer a sentencing factor. Against, in three moves the machine cannot answer. Desert: punishment answers this act by this person; an actuarial score punishes a predicted future, and punishing the predicted is the innocence constraint's cousin — some part of the sentence lands on acts never committed. Fairness: strip the forbidden column and the zip code carries it back in; the like-cases requirement does not care that the special pleading was automated. Opacity: sentences owe reasons, and "the model said so" fails the requirement set in the course's first week — while the responsibility gap asks who answers for the outlier: vendor, county, judge, or nobody, which is the problem. The values line, named for the record: whether punishment answers the past or manages the future — a dispute two centuries older than the software, now running at scale, minus the explanations. The file's finding: the county did not buy an answer; the county bought its oldest question, laminated.