Semantic Drafting: Preserving Multidimensional Expert Judgment in LLM Output through Layered Constraints

This paper proposes Semantic Drafting (SD), a framework for preserving multidimensional expert judgment in LLM outputs through layered semantic constraints. SD organizes reasoning through L1-in, L4, L3, L2, and L1-out, and examines how upper-level constraints propagate into lower-layer generation. Through SD prompt experiments and practitioner evaluation in landscape-related tasks, the paper identifies failure modes such as bleaching, drift, forgetting, and structural collapse. It introduces L4 Constraint Quality (LCQ), Authority Check, and an SD quality management cycle to diagnose and correct failures in constraint propagation. SD is positioned not as a replacement for expert judgment, but as a quality-managed reasoning design framework for making AI-assisted judgment more observable and improvable.