Governance

A governance model AI teams can actually operate.

Most AI governance frameworks live in slide decks. DataVibe enforces them at runtime. Every outbound payload passes through three layers, policy, approval, audit, with graduated trust to keep operators out of the loop when it is safe to do so.

1. Policy versions

Each workspace has named policies (e.g. "outbound-default", "support-replies"). A policy has versions. Only one version is published at a time. Operators can preview a new version against historical payloads before promoting it.

2. Approval queue

Payloads that fail policy or land in a WARN tier route to a human approval queue scoped to a workspace role (Reviewer, Admin, Owner). Approvers see severity, a payload preview, the rule that fired, and the recommended action.

3. Trust graduation

Each action type carries an autonomy trust state, manual, semi-autonomous, or autonomous. Operators graduate an action type out of the queue only after a configurable window of clean reviews. Any policy violation collapses trust back to manual.

4. Immutable audit

Every gate decision lands in an append-only audit row that captures the payload, the policy version that ran, the operator (if any) and the dispatch result. Exports to CSV/JSON for compliance handoff.

5. Roles

• Owner. Workspace billing, settings, members, and content.
• Admin. Members, policies, and content; no billing.
• Reviewer. Approval queue + audit; read-only policy access.
• Developer. API keys + integrations + docs.
• Billing. Invoices and plan changes; no operational surface.