Every generated answer carries citation chips. There are four kinds:
| Chip | Means |
|---|---|
| ⚙ Control | e.g. SEC-01 · implemented — the control and its current status |
| 🧪 Evidence | e.g. tested 2/2 passing · 7/25/2026 — pass counts from your latest run, with date |
| 📄 Policy | e.g. AI Risk Management Policy v1 — an approved policy covering that control |
| ★ Answer bank | a previously verified answer was reused |
Why they're trustworthy
Citations are assembled deterministically from your database — not written by the language model. The model's only role is choosing which of your scoped control IDs are relevant and phrasing the prose. Control statuses, test counts, dates, and policy versions are looked up directly.
So a citation can't be invented, and it can't claim a policy exists that you haven't approved.
Why this matters commercially
A security reviewer who catches one unverifiable AI-written answer stops trusting the entire submission — and often the vendor. Citations turn "trust our summary" into "here's the control, here's the test result, here's the date, check it yourself."
When there are no chips
No chips means nothing in your data supported an answer. Those get flagged needs review rather than dressed up. That's the honest failure mode, and it's why the flag exists.