Part 1 of 4 · 1 min read
Overview
Production is where you find the failures your eval set missed. Interviewers ask "how do you improve an agent after launch?" A weak answer is "look at thumbs-down". A strong one explains bias, triage into eval cases, expert labels and a controlled rollout.
About the example, and what is real
Real: the simulation code (response bias, winner's curse, proxy optimisation, flag quality, round-robin triage) and the 13 tests in code/mini-feedback. Assumed: unhappy users give feedback at 6 percent and happy ones at 1.5 percent; variance among prompts; users' thumbs weight agreeableness by 0 to 0.7; a downvote and regeneration rate for failing and passing answers. Not real: no real users, no real feedback data; the agreeableness trade-off is a modelling choice, so the size of the damage is an illustration only.