For one automation (or AI-drafted recurring output), run a propose-only week: it drafts, you review and send. Tally its errors. You're done when you can name its most common mistake — and decide whether it's earned more rope.
How-to · AUT · AUT.2
Week one: propose, never execute
New automations should run in propose-only mode first — outputs reviewed by a human — before earning execution rights.
Worked example
Every automation deserves a probation period. Week one: it proposes, you execute. You're not slowing it down; you're collecting its error patterns while the blast radius is zero. Execution rights are earned with a clean track record, not granted at setup.