Abstract
Legal question answering is usually evaluated as if the law were static. In Kazakhstan, it changed twice within one year: a new Constitution replaced the 1995 text on 1 July 2026, and a new Tax Code replaced its predecessor on 1 January 2026, both reusing article numbers for different norms. We release a dated corpus of 41 Kazakhstani acts parsed to article level in both official languages (12,235 Russian and 11,973 Kazakh articles) and a benchmark of 150 question-answer pairs in seven layers, two of which score declining to answer and asking for a date. Two practising lawyers confirmed every pair. Comparing a no-retrieval baseline, retrieval-augmented generation and a tool-using agent, we find that retrieval rather than generation accounts for most citation error, measured against each pair’s source articles: given the gold article, the model cites it in 92% of cases, and retrieval alone reduces citations of non-existent norms from 31% to zero. The agent adds nothing on this measure and abstains far less often. Dense retrieval improves citation recall but nearly doubles the repealed norms presented as current; date-aware ranking removes them for every retriever, although under dense retrieval it also removes three historically required citations. The retrieval findings are reproduced on two open models.