Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law

We identify and quantify a previously understudied failure mode of retrieval-augmented LLMs in the legal domain: temporal misgrounding, the systematic retrieval and citation of the currently in-force version of a legal article when the question implicitly requires an earlier or future version. We argue that legal question answering should be reframed as a temporally-indexed retrieval problem rather than a static one. To support this claim, we introduce FiscalQA Pro, a benchmark for grounded reasoning in French tax law that explicitly evaluates temporal grounding. It pairs a temporally-versioned corpus of 32,436 article-versions of the French General Tax Code and Book of Tax Procedures (93 years, 1938–2031) with a temporal-reasoning track (R3) of 35 expert-curated, all-frontier-hard questions, scored deterministically via KARLBench-style nuggets rather than LLM-as-judge. Our killer experiment measures the cost of temporal misgrounding across four frontier models (k=35): frontier LLMs reach only ~7% exact-answer accuracy without retrieval (mean strict over repeated draws; never above 14% on a single run) and 11% with RAG over a static, current-version corpus—which retrieves the date-applicable version 0% of the time. Conditioning retrieval on the query date closes most of the gap: 87% with oracle version selection, and 83% with an end-to-end production retriever feeding the top-5 date-applicable versions (gold version present 94% of the time). We additionally release a version-aware jurisprudence dataset of 69,208 citation links between 32,034 court decisions and the date-applicable article version (98–99% link-level precision), and all corpus, benchmark, and pipeline artifacts.