Journal
Guardrails that belong on the brief, not on Friday’s curiosity
A guardrail is a metric you are willing to stop a test for, written down before assignment. It is not a metric you found while browsing because the primary number looked awkward.
For a checkout experiment we usually want refunds or cancellations in the same window, and a payment-failure rate if the treatment touches the payment step. For onboarding, completion without a spike in fraud referrals. Three guardrails are enough. Twelve is a way to keep every stakeholder quiet and no one accountable.
The brief should say what movement would pause the test. “Watch refunds” is not a rule. “If refund rate on the treatment is clearly worse after the same exposure, we hold” is a rule a Friday meeting can use.
Curiosity metrics can still be listed as exploratory. They simply do not get a veto unless you amend the brief in writing — which, mid-test, we will name as a change, not as original intent.