DEX‐LL: Data‐induced Decision EXpert with large language models

Abstract Operational decision support systems often rely on ordered risk states rather than calibrated probabilities and therefore require transparency and consistency. This paper introduces DEX‐LL , a data‐inductive extension of the Decision EXpert (DEX) qualitative decision‐support framework that constructs hierarchical ordinal decision models with explicit, monotone decision tables. Intermediate decision concepts are learned by aggregating small subsets of categorical attributes, with monotonicity enforced as a semantic requirement via isotonic projection on discrete ordinal grids. To stabilize rule induction under rare events, conservative uncertainty‐adjusted scoring based on Wilson lower confidence bounds is employed. Large language models are used only to propose candidate attribute groupings for intermediate concepts, while all rule induction and validation remain fully data‐driven and deterministic. Evaluation on six seasons of ski resort operational data shows that DEX‐LL yields interpretable, auditable, and action‐consistent risk stratifications that remain competitive with standard predictive models when evaluated on an operationally aligned ordinal decision scale.