Human-in-the-loop design that isn't rubber stamping

My automation does {{describe}} and I want a human approving before {{action}}. Volume: {{per day}}. Cost of a bad approval: {{consequence}}.

Design the review step so it's real:
1. Define what the reviewer is actually verifying — the two or three things that would make this wrong. Not "check it looks right".
2. Design what's presented: the output, plus the specific evidence needed to verify it, plus what the automation is uncertain about. Everything else hidden.
3. Fight rubber stamping: what fraction will be approved without reading at this volume? Propose the mechanism that keeps attention — sampling, forcing a field, flagging only the uncertain ones for review, or batching by risk.
4. Decide what genuinely doesn't need review, so the reviewer's attention goes where it matters.
5. Define the reject path: what happens, what feedback is captured, and how it improves the system rather than just fixing the instance.
6. Set the SLA and the behaviour when nobody reviews in time — queue, auto-reject, or escalate. Never auto-approve on timeout.
7. Log the reviews so I can measure whether the human is catching anything. If they never catch anything, the gate is theatre.

Be honest about where a human gate adds no safety.