Healthcare AI earns its 95% accuracy scores against clinical-language test sets — structured queries, precise symptom descriptions, and scenarios written by people who know what they are looking for. The patients using these tools at midnight are doing none of those things. When real-world conversational queries replace benchmark prompts, accuracy collapses to 33%. That is not a performance edge case; it is the median patient experience. Deploying a system with that gap into after-hours care — where the human alternative has explicitly been removed — means the accuracy number on…