blind-evaluation

Design and run a head-to-head comparison whose result you can actually believe — blinding, balanced position assignment, a control that could plausibly have won, and honest reading of a small sample. Use this whenever you are about to claim one approach beats another: A/B-ing two prompts or configs, comparing a new pipeline against the old one, benchmarking a library or model against an alternative, testing whether a change actually helped, or judging which of two outputs is better. Also use it when someone hands you a comparison result and asks whether it holds up, when a result looks suspiciously clean, or when you notice you are about to declare a winner after looking at the outputs unblinded. Not for checking that your own work is finished (use verification-before-completion) and not for wording the claim once evidence is in (use falsifiable-claims).