Answering questions grounded in normative documents often requires information beyond a single passage: whether the retrieved document is the currently effective version, whether it applies to the relevant jurisdiction, entity (e.g., institution or applicant) and date, and whether each normative claim can be traced back to its source text. Hosted retrieval services have dramatically reduced the engineering effort needed to build an initial system over such corpora, making “upload the documents and ask” a common default. We evaluated this default on roughly 73,000 candidate normative documents in a production deployment, using a stratified sample of 200 questions from our published benchmark, each with a gold source document; the sampling rule does not read system outputs or scores. We compared the hosted service with a governed system that resolves version and scope through explicit rules before generation. The governed system achieved an overall score of 97.7, while the hosted service scored 88.1, a gap of 9.6 points computed from unrounded means. The governed configuration has been offered as a commercial product since January 2026, serving 1,126 registered users, including customers such as Zhipu AI and Lecheng Health. By mid‑April 2026 it reached roughly 100,000 calls per workday.
Review