In December 2025 the Supreme Court of Korea adopted a staged Judicial AI Roadmap. Sentencing-system improvement and recidivism assessment sit among its declared tasks. A public-sector institution has thus committed to governing adjudication, a liberty-bearing state function, with AI, and the questions studied internationally under the heading of Legal Judgment Prediction (LJP) have become immediately practical for Korea. This article maps Korea’s LJP literature in a scoping review. Reading that corpus against the global field, it claims that the field’s defining feature is a normative-empirical disconnect. Its normative (legal) and empirical (technical) literatures have developed separately, with little cross-engagement between them, in that no study in either literature is co-authored by an engineer and a jurist, and no normative work evaluates a domestic model. This article makes two contributions from this diagnosis. First, it argues that engineer-jurist collaboration must tie the technical design of Korea’s LJP to the culpability principle, reason-giving, and judicial independence, treating these constitutional values as design constraints rather than ex post safeguards, and proposes that governed data access and interpretable baselines with legally legible feature attribution (e.g., SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME)) precede deployment. Second, it re-conceives the prediction target of LJP datasets for Korean criminal procedure as three distinct stages, which are guilt determination, applicable provision, and sentence, showing that task definition is itself a legally consequential design decision.