Privacy
What we store, field by field.
llm11 is a proxy. Your request passes through our infrastructure so it can be routed and checked, which means we see it in flight. What happens after that is the part worth being specific about.
Stored on every request
- Which model answered, and which model your project declared as its baseline
- Prompt and completion token counts
- Cost, baseline cost, verification cost and the resulting saving
- Which checks ran, whether each passed, and its confidence
- Whether the request escalated, and which check caused it
- Task type and criticality, as triage classified them
- Latency, end to end and for the triage call alone
Not stored, by default
- The text of your prompt
- The text of the model's answer
- The context you passed for the groundedness check
- Anything your users typed
If you switch retention on
Retention is off by default and is set per project, not per account. Switching it on stores the prompt, the response and the context for a window you choose, up to 90 days, after which the sample is deleted. It exists for one reason: when a groundedness check fails and you want to see what it was actually looking at.
Retained content is never used to improve routing or verification for any other customer, and it is never used to train anything.
Who else sees the request
The model provider that answers it, reached through Vercel AI Gateway or OpenRouter, under their own data terms. Where a triage backend is configured, that vendor sees the request too. Those are the only third parties in the path, and each one is named on the receipt for the request that reached it.
A working draft, published so the posture is stated rather than assumed. It will be replaced with counsel-reviewed terms before this product takes production traffic beyond its design-partner group. Questions to hello@llm11.com, or read the terms.