Design an experiment that can actually conclude

I want to test {{change}} on {{surface}}. Baseline: {{metric and current value}}, traffic {{volume per week}}.

Design the test:
1. State the hypothesis as a mechanism, not a hope: "changing X will cause Y because Z".
2. Primary metric (one), guardrail metrics, and the metrics I should look at but not decide on.
3. Minimum detectable effect worth shipping — reason from the business value, not from convention. Then compute the required sample and duration at 80% and 90% power. Show the numbers.
4. Tell me honestly if my traffic can't detect a realistic effect in a reasonable time, and what to do instead (bigger swing, upstream metric, switchback, or don't test).
5. Randomisation unit and the interference risks with it.
6. Pre-register the decision rule: what result ships, what kills it, what's inconclusive — written before we look.
7. The three ways this test will be misread when the numbers come in.

No peeking advice, no 'run it and see'.