Classifying a grounded prompt's sensitivity based on what data it contains and what it's being asked to do with it — distinguishing a routine QBR summary from a prompt that would export unmasked financial or health data — to feed the routing decision downstream.