The rewrite dropped a logical operator in the user's question. “Not eligible” is the target set, not a decorative word to remove for cleaner search. A dense retriever may still find semantically close eligibility text, and an LLM can turn that into a fluent answer about the opposite group. Research on domain-specific query rewriting argues that rewriting needs to reason about the information need rather than just transform surface text. That paper is not a claim about this particular policy corpus. The local evidence is the original query, rewrite, candidate list and answer.

I would preserve an intent representation alongside any expanded query: entity contractor, benefit travel allowance, requested class excluded, effective date and jurisdiction if supplied. Run exact lexical searches for exception and exclusion language, plus semantic candidates. Do not translate natural-language “not” directly into a broad Boolean NOT eligible filter without checking the document structure. The relevant exception may appear in a paragraph that also contains the word “eligible,” and a Boolean exclusion could remove it. Use retrieval to find the governing rule and exceptions, then apply the rule to named people only if their attributes are available and authorized.

Evaluate the rewrite as a separate stage. Compare original and rewritten queries against a small set of contrast pairs: “eligible” versus “not eligible,” “before” versus “after,” “with approval” versus “without approval,” and policy versions. Inspect first-stage recall for the exception source, then whether reranking and answer synthesis keep the exclusion. If the source does not enumerate all contractors, answer the rule rather than fabricate a list of people.

The interviewer may argue that the final generator still sees the original user question. It does, but it cannot cite an exception the retriever never supplied, and can easily rationalize the wrong broad policy. The query rewrite removed a version number. Why did RAG still look healthy? covers a version number lost in a rewrite. The retriever found the exception. Why did the reranker put it last? covers an exception that retrieval found but reranking demoted. This one loses the negation before retrieval and points the whole evidence search at the wrong population.