The Silent No: Measuring and Eliminating Silent False Negatives in AI Retrieval over Regulatory Text

Retrieval tools for pharmaceutical regulatory text almost all ship the same architecture: semantic chunking, vector embedding, and top-k retrieval feeding a language model. The failure that matters most in a regulated setting is rarely measured. When such a system returns nothing for a question whose answer is plainly present in the governing document, the empty result reads as a clean "nothing applies here." We call this the silent false negative. We built a frozen, hash-pinned gold set over a public-domain corpus: 21 CFR Part 211 ingested from the official eCFR as the oracle, plus three final FDA drug cGMP guidances (Aseptic Processing 2004, Data Integrity and cGMP 2018, Process Validation 2011). Gold items were generated from the documents' own structure rather than hand-picked. Three systems were compared across three model families (llama-3.3-70b, gpt-4o-mini, mistral-small-2603): System A, answer-time vector retrieval; System B, a build-time-validated structured gate with deterministic lookup and abstain-with-pointer; and System B2, System B plus a field-level abstention gate. On 258 pooled answerable items, the silent-false-negative rate was 63% (95% CI 57 to 69) for vector retrieval and 5% (95% CI 2 to 8) for the structured gate, with correctness rising from 24% to 95% and wrong-positive answers falling from 12% to 0%. The paired McNemar exact test for A versus B2 gave p = 4.2e-39. The direction held for every model family and all three guidances, where the structured tier reached 0% silent false negatives on each document. The field-level gate lowered hallucination on unanswerable items from 19% to 8% while holding the 5% silent-false-negative rate and 95% correctness. Build-time-validated structured extraction relocates retrieval error from a silent answer-time miss into a bounded, reviewable build step. The benefit appears architectural and model-invariant across the three families tested. One residual limitation, a word-overlap imprecision in the field gate, is reported rather than tuned away. This deposit contains the preprint, the pre-registration, the frozen gold set, the public-domain corpus (SHA-256 pinned), the full pipeline code, the canonical scored run, the figures, and the reference-integrity record. Human-reviewer sign-off data is deliberately excluded. Code is licensed MIT; paper, data, and documentation CC-BY-4.0.