Toward Composable Intelligence: A Research Roadmap for Jev

Jev maps textual state to categorical choices, ordinal scores, and Boolean probabilities. We organize current methods and systems into four functional families. Experiments with TF–IDF, MiniLM, BM25, and a cross-encoder on BANKING77, CLINC150, and SciFact illustrate component interactions: improved in-domain calibration can accompany higher out-of-scope false acceptance, while greater candidate recall can coexist with lower observed ranking quality. We develop a roadmap for reusable decision programs that preserve the meaning of criteria, evidence, uncertainty, and action conditions as their components and tasks change.