pdca-do

Design a small controlled pilot with defined scope, duration, and rollback conditions, capture a pilot baseline before changing anything, then implement the plan actions sequentially while collecting the same metrics as the baseline. Document all deviations and observations in the pdca-do.md artifact.

A completed pdca-plan.md with a numeric baseline, a SMART objective, a formulated improvement hypothesis, and an action plan with owners and dates. Without a baseline, the Check phase cannot compare results, and without a rollback plan the pilot is unsafe for critical systems.

Changing multiple variables simultaneously makes it impossible to attribute results to a specific action. If metrics improve or degrade, you cannot identify the cause, so the hypothesis cannot be confirmed or rejected in the Check phase.

Stop and revert when degradation in the main metric exceeds the predefined threshold, when an unforeseen critical production error occurs, or when the maximum pilot duration passes with no positive signal. These stop criteria must be defined before the pilot begins.

Judging whether a pilot is representative requires domain knowledge, and processes with seasonal variability need pilots covering at least one full cycle. If the pilot cannot be isolated from production, external conditions that could contaminate the data must be documented.