# AI guardrail review

A review aid, not a certification.

- [ ] Every published claim has an identified source.
- [ ] Missing or conflicting information is visible, not guessed.
- [ ] The system retrieves only information the user is allowed to access.
- [ ] External content is treated as data, not authority to change the task.
- [ ] Read actions and write actions have explicit boundaries.
- [ ] A named person approves consequential actions before they happen.
- [ ] Outputs are checked against the brief, format, and brand standards.
- [ ] Logs retain the decision, status, and owner without unnecessary sensitive data.
- [ ] A planted failure has demonstrated that the check actually blocks the work.

Reviewer: 
Date: 
Follow-up owner: