Guidelines move on, answers don't: with a strong model, retrieving several ASCO guideline versions made stale answers about four times as frequent as answering without retrieval

Clinical question answering has a measurable failure mode: an answer that is faithful to a real guideline but to a superseded version of it, so a citation check still passes. A hallucinated claim has no source and can be caught by checking the reference; a stale claim has a legitimate source, and the source itself gives no warning that it has been replaced. Across the 288 ASCOBench test turns, oncologists had to revise or rewrite about one in four version-sensitive answers drafted by a frontier model, while factual answers were accepted 92.3% of the time.

The system verifies an answer before display: the factual verifier tests each claim against the retrieved passages without looking at the citations the answer gives, and labels the answer verified, partially verified or unverified. The temporal verifier checks that the most recent guideline version takes precedence, and a claim may be flagged as outdated only with a verbatim quotation from the superseding passage, which stops the model from relying on what it remembers of guideline history. On the 288 unique test turns, SentryLine gave incorrect answers at 4.2%, 4.5% and 5.2% with Gemini 3.1 Pro, DeepSeek-R1 and Llama 3.3 70B, against baseline ranges of 13.9-18.8%, 9.7-12.5% and 17.7-22.9%.

The evaluation was run on ASCOBench only, covers ASCO breast and prostate cancer guidelines, and uses model-generated, clinician-reviewed conversations rather than real encounters; SentryLine addresses the case in which a guideline has already declared a change.

Stale ratings come from a judge rubric that could not be validated against the oncologists' labels because only 2 of 88 drift turns were negative, so they should be read as exploratory; the verifier ablations were small and showed a clear effect only for DeepSeek-R1, with no detectable effect under Llama 3.3; the seven-dimension annotation covers 54 of 96 test conversations, making the drift subset a lower bound; recognizing a change before a guideline declares it remains an open problem.