How we check section retrieval
What was measured
The downloadable plain-language-retrieval-v1 benchmark contains 57 everyday English and Pidgin-style questions and expected sections. The recorded 12 October 2026 run used the original 55 cases; its result reports a 0.98 top-one hit rate and 1.0 top-five hit rate after adding the topic guide. This is a retained retrieval result, not a new test of the current deployment.
What the result does not establish
A retrieval hit means an expected section appeared in the measured search position. It does not measure the correctness of generated explanations, legal advice quality, completeness of the corpus, current-law coverage or performance on questions outside the test set. The questions are a small curated set, not a representative sample of all users.
Reproducible scope
The versioned test file records each question and its expected sections. A repeat test should record the deployed revision, published corpus release, retrieval configuration, date and per-question results. Compare the same dataset before reporting a change. Source wording and later amendments require separate verification.