Answers are grounded in this paper and cite their sources.
Defect disputes in Korean apartment housing are adjudicated under a ministerial notice that defines defect criteria and an enforcement decree that fixes defect liability periods (DLPs) by work type. Large language models (LLMs) can support this work only if grounded in these regulations. This study proposes an LLM agent that extracts structured facts from a case record, retrieves articles of the notice, applies its numeric thresholds through an executable rule engine, and determines the DLP with a deterministic lookup tool. The agent was evaluated with two commercial LLMs on 262 adjudicated cases from three official casebooks, with the 2026 casebook (128 cases) as the main evaluation set, and on 67 contrast cases that vary only a measured value around a threshold. Access to the statutory DLP table, in the prompt or via lookup, raised DLP accuracy from 0.44–0.64 to 0.88–0.89, whereas article retrieval alone gave little or no gain. Retrieval raised article-citation F1 from about 0.06 to 0.66–0.71. On contrast cases, retrieval-augmented accuracy differed by model (0.84 vs. 0.97), while the agent reached 0.97–1.00 with both. Statutory tables and executable rules make DLP and threshold judgments reliable, while work-type classification remains the main error source.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: