Architecture boundaries

RaiCentral owns: agent registry and versions, runtime identities and credentials, policy evaluation, human-approval workflow for governed actions, decision receipts, and evidence linked to reported results.

Your application owns: prompts, personas, business logic, and the actual side effects (refunds, ERP writes, MCP tool calls, etc.). RaiCentral returns a decision; your code still performs or withholds the action.

Trust rules: clients must not self-assert allowlists or risk outcomes. Authorize before side effects. Fail closed when the plane is unavailable if the action is consequential. Approvals are bound to the specific action request.

Stable external contract: integrate through the versioned /sdk/v1 HTTP API (or the official SDKs that wrap it). Treat internal module APIs as unstable.

Coverage depends on integration: only actions your application routes through authorize are governed. RaiCentral does not magically wrap every network call.