Overview
The hypothesis, change, result, decision, and notes recorded for every run.
Use small changes wherever possible. When several variables change at once, the log cannot clearly identify what caused the result. Small, reversible experiments create cleaner learning and make regressions easier to undo.
The same structure can be applied beyond code performance when the three requirements exist. A prompt can be changed and evaluated against a repeated support task. A page can be changed and evaluated against a conversion measure. An outreach message can be changed and evaluated against response data. These uses require enough observations to make the assessment meaningful and a loop fast enough to justify automation.