This paper proposes a procedural institutional architecture for handling collective human judgment onmorally or legally contested questions, in which a language model performs task decomposition andcomplexity isolation while final substantive judgment remains with human reviewers (e.g., jurors, expertpanels, or crowdsourced volunteers). The central design problem is not whether AI can adjudicate valueconflicts, but how an institution can identify, before aggregation, which forms of reviewer disagreementreflect genuine value pluralism and which reflect measurement artifacts, and route each accordingly.Drawing on Kaplow’s rules-versus-standards distinction and Dancy’s situationist measure to bound whatcan be decomposed into rule-governed sub-tasks, and on List and Pettit’s judgment-aggregation theoryand Fogelin’s theory of deep disagreement to diagnose structural risks in recombining sub-judgments,the paper proposes a pre-aggregation disagreement-structure analysis layer: a procedural component,triggered jointly by salience and degree of disagreement, that routes high-risk aggregation outcomes to ared-team and human-adjudication process rather than resolving them algorithmically. The paperpositions this contribution against three adjacent literatures — AI-assisted deliberation, moral uncertaintyand value pluralism, and disagreement-preservation approaches in machine-learning annotation(including jury learning) — and argues that its distinct contribution is procedural: locating disagreementdiagnosis before, rather than during or after, aggregation into an institutional decision. The architecture istheoretical: it has not been validated against real case data, and its proposed thresholds require futureempirical calibration.
Hanoi Towerz (Wed,) studied this question.