Skip to main content
The full Commandments live at inviolet.ai/commandments with explanations and product-trace mappings. The summary:
  1. Every action declares a purpose. No purpose, no action.
  2. Purpose lives outside the agent. The system records and enforces it; the agent doesn’t get to decide what it just did.
  3. Identity is necessary but not sufficient. Who is asking matters. Why they’re asking matters more.
  4. Default to deny. New purposes are observed before they’re allowed.
  5. Allow narrow, deny broad. Approve elements one at a time, not whole tables.
  6. Bound credentials by purpose, not by user. Standing credentials are a liability; just-in-time credentials with TTLs are the answer.
  7. Audit everything. Every allow, every deny, every approval is queryable.
  8. Humans approve the categories. Agents stay within them. Don’t make humans approve every query; make them shape the policy.
  9. Anomaly is a question, not an answer. A first-time pattern is a conversation between operator and agent — not an automatic block.
  10. The right hand always knows what the left hand is doing. Single intent-token claim format across IdP, gateway, vault, and DB proxy.

How these map to Inviolet

Each commandment traces to a specific phase of the build. See the Inviolet phase plan for the long-form mapping.