We A/B test model outputs before shipping — a cheap trick that works

Every prompt change gets a shadow run: same input, old and new config, human review on a sample. Catches regressions the evals miss. Takes one extra table and a cron job. Best $0 investment we've made this quarter.