The interesting failure here is not that a model can invent a plausible citation, but that the surrounding workflow let it pass as verified. A citation is a verifiable claim about another document, so it should be checked by retrieval and matching against the source, not by asking the same generation process whether it sounds right. In my experience, the fix is boring and effective: keep generated text clearly marked as a draft, require a separate check for title, date, and quoted content, and block submission when any check is missing. Without that separation, confident-sounding references will keep slipping through precisely because they look finished.