Divergent precedent

Decisions get made one case at a time and kept as a record that later decisions draw on. Two genuinely different disciplines live under that substrate: writing each outcome as narrowly as possible, so parties who disagree about the general principle can still sign the same specific holding (the legal idea of an incompletely theorized agreement), and arguing a new case out of the record by citing the closest past cases, distinguishing the ones cited against you, and manufacturing hypothetical variants to find where a position breaks. These pull in opposite directions on one point, and the regime should show that rather than hide it: the first deliberately leaves the principle unsettled, the second is a search for the principle's boundary, and a system can do either without the other. A benchmark here asks whether language models can tell when one Supreme Court decision overrules an earlier one -- a capability any accumulating case record presupposes, though neither of the two disciplines above actually supplies it.

Every rater's label is kept as that person's own judgment, often alongside information about who they are, and models are trained and scored on reproducing the whole spread. Datasets and shared tasks here rate each item many times over and grade a system on how well it predicts either the population's distribution of views or a named individual's view, while analysis in the same vein shows that standard Bradley-Terry reward modeling and LLM-as-judge evaluation discard exactly that spread. What this family does not do is worth stating plainly: it preserves divergence in the record, but nothing in it cites an earlier judgment or lets one decision bear on a later one, so it supplies raw material a precedent system would need without any precedent machinery of its own.

A community argues over concrete cases and changes the thing that governs it. The governed thing differs across the family and the difference is real: PolicyCraft grows a written policy document from submitted real cases, tracking which clauses each case grounds and surfacing cases that contradict the policy as it stands, while Botender changes the deployed behavior of a group's bot, using generated interaction scenarios as provocations rather than waiting for cases to arrive. What holds them together is the loop rather than the output format -- ordinary members deliberate over particular cases, and a shared governing artifact, document or agent, is edited as a result. Sitting alongside the tools is a measurement framework that splits a group's uncertainty on moderation cases into ambiguity, which a clearer description of the case can reduce, and genuine disagreement, which it cannot; it belongs here because it tells builders of these systems which arguments are worth staging, not because either system performs that split.