Citation-grounded modeling links a prediction, retrieval result, or review-support signal to the documentary evidence that made it possible. In legal analytics, citation and phrase histories help represent precedent, judicial alignment, and panel behavior. In peer-review support, citation-grounded phrase profiles and structured review evidence help estimate agreement, expertise coverage, and uncertainty. In enterprise retrieval, semantic candidates must be joined to structured records, access constraints, and chronology before they can support a decision. This survey reviews citation-grounded modeling across legal, review, and retrieval systems from a 2023 perspective. We synthesize hybrid semantic-relational retrieval, long-term forecasting, citation-grounded reviewer profiles, agreement-gated semi-supervised learning, structured extraction, GPU-parallel optimization, and prior archive work on judicial analytics, reviewer agreement, automated peer review, semantic integration, self-ensembling, and gated pseudo-labeling. The survey defines a five-layer evidence model covering source capture, citation and phrase alignment, profile construction, constrained retrieval, and confidence-gated publication. A comparative coding and analytical study show that the strongest systems separate broad evidence discovery from narrow output eligibility. The result is a conservative design rule: legal, review, and retrieval systems should expose citation support, cutoff dates, missing evidence, and confidence before their outputs are used.
Kumar et al. (Thu,) studied this question.