Governance
Data decides where AI is allowed to run.
Each classification sets whether work may leave the device, which providers are approved, and when a person must review the output. When the rules can't be met, the system fails closed.
Each classification sets whether work may leave the device, which providers are approved, and when a person must review the output. When the rules can't be met, the system fails closed.
Read-only on this public site. These are the policies every request here is routed under, and the access matrix is computed live by the same function the router uses. Editing works on a local install and is rejected server-side here.
The more sensitive the data, the fewer places it may go. Computed from policy v1.
Safe to share
Company only
Client or sensitive data
Regulated or personal
If restricted data must stay local and no local model is available, the request is blocked rather than sent to the cloud. Failing closed is the policy working, not an error.
Version 1. Applies to every routing decision, comparison run and fallback.
| Classification | Cloud access | Approved cloud providers | Prefer local | Human review |
|---|---|---|---|---|
| Public | ||||
| Internal | ||||
| Confidential | ||||
| Restricted |
Which model may receive which data, computed live with the same function the router uses.
| Data | Gemma 4 12B | Qwen 3.5 9B | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|---|---|---|
| Public | Allowed | Allowed | Allowed | Allowed | Allowed |
| Internal | Allowed | Allowed | Allowed | Allowed | Allowed |
| Confidential | Allowed | Allowed | With exception | With exception | With exception |
| Restricted | Allowed | Allowed | Blocked | Blocked | Blocked |
When an output waits for a person before anyone uses it.
Enforced in code, so no setting can turn them off.
Changes take effect on the next routing decision. Each request's decision snapshot records the policy version that produced it.