Concepts

Session lifecycle (typical): heartbeat → start session → observe input → authorize action → execute only on ALLOW → report execution → optional observe output / metering → end session.

Authorize before act. Report after ALLOW. Post-hoc logging is not enforcement — if your application never calls authorize, RaiCentral cannot govern that action.

Decision meanings: ALLOW (allowed=true) — run the side effect, then report; DENY — do not run; REQUIRE_APPROVAL — hold for a named reviewer; the approval is bound to the specific action and parameters; MONITOR — do not treat as permission to execute unless your integration policy explicitly says otherwise.

Registry and identity: register AI systems and agents with owners and versions; issue scoped runtime credentials (agent keys) for the identities that call the SDK. Credentials should be short-lived where possible and never committed to source control.

Policies and approvals: policy evaluation runs at authorize time. When a rule requires a human, the application waits; after approval, authorize again or follow the approval workflow your deployment exposes. Approvals are action-bound — they apply to the requested action and parameters, not a blanket grant.

Evidence: every decision produces a receipt. Reported execution results (and optional spend/metering) link back to that decision so reviewers can follow what was requested, decided, and reported.