# Field Guide Chapter DSL Sample
Field Guide
How to read AI trust boundaries
AI systems fail at the points where trust changes: user to app, app to gateway, gateway to retrieval, gateway to model provider, gateway to tools, and gateway to logs.
Quote
“The boundary is where the control belongs.”
Note
Practical rule
When an AI component receives new data, authority, or context, ask what trust assumption just changed.
Checklist
Boundary review checklist
✓Identify the actor.
✓Identify the data.
✓Identify the authority.
✓Identify the system receiving it.
✓Identify the control.
✓Identify the log.
✓Identify the owner.
Common AI boundaries
| Boundary | Main risk | Control |
|---|---|---|
| User to prompt | untrusted input | input policy |
| Prompt to retrieval | unauthorized context | source authorization |
| Gateway to provider | third-party data exposure | minimization |
| Gateway to tools | state change | permission matrix |
| Tools to logs | weak reconstruction | audit trail |
Decision · implemented
Reader decision
If you cannot name the control at a boundary, the system is not ready for serious review.