You're evaluating a new product feature and need to decide whether an experiment will give a trustworthy read. You want a clear primary metric, a sensible minimum detectable effect, and guardrails so a local win does not hide broader harm.
How do you approach A/B testing for a product feature?