Protect data

Policy rules

Rules decide which models may answer a request. They're checked top to bottom; the first rule that matches decides. If none matches, the request is refused.

How rules work

  • A rule matches on what's in the request (data classes, task, team, workload) and names where it may go.
  • Targets can name specific models, or describe them: "on-prem, then private cloud", "processing in the EU", "provider holds ZDR or NO_TRAINING", "good at classification, cheapest first". Described targets resolve live, so they keep working as providers and models change.
  • The key's frameworks then narrow the result further, for example to providers with a signed BAA.
  • Rules can also deny outright, and the last rule is usually a catch-all.

Default rules

New organisations start with:

#RuleEffect
1health-data-on-prem-firstHealth data goes to on-prem models first.
2government-data-us-onlyGovernment-controlled data stays in the United States.
3personal-data-off-public-internetPersonal data doesn't go to public APIs.
4confidential-private-cloudConfidential data stays in your private cloud.
5everything-else-requested-modelEverything else goes to the model the app asked for.

Policy library

Under Policies → Policy library (opens in a new tab) there are 22 ready-made rules in five groups: Data protection, Data residency, Vendors & contracts, Cost & quality, and Catch-all. Each is added where it belongs: blocks at the top, catch-alls at the bottom, everything else just above your first catch-all.

Test before you change

The Playground (opens in a new tab) shows the decision for any message and profile without calling a model. Every policy change is snapshotted; roll back from the policy history if a change misbehaves.