Instrument an A/B test so the result is actually trustworthy

Instrument an A/B test so the result is actually trustworthy
The setup checklist before you read a single p-value.

1 — Assign variants server-side and persist the assignment to the user record.
Why: client-side flicker and re-randomization on reload silently corrupt your splits.

2 — Fire an 'exposure' event only when the variant actually renders.
Why: counting assignment instead of exposure inflates your denominator with users who never saw it.

3 — Lock your primary metric and sample size before launch.
Why: deciding the metric after seeing data is how noise becomes a 'win'.

4 — Add a guardrail metric you'd never want to harm.
Why: a checkout lift that tanks refunds isn't a win; instrument the thing you're afraid to break.

That's the stack for this week. Forward to a teammate.