From an agent's request to an authorised, evidenced action

From an AI system in the inventory to an authorised, evidenced action.

Two layers, one record

Both layers share the same AI systems, controls and evidence. A control mapped to CPS 230 in the registry is the same control a gate policy enforces at runtime, and the decision receipt it produces becomes evidence for that control.

Step by step

1. 1
   The AI system is known and governed

   It is in the inventory with an owner, a risk tier and the controls that apply.

2. 2
   Context becomes governed state

   Facts the agent relies on are recorded as typed, versioned state items with their source.

3. 3
   Conflicts are detected

   When sources disagree, a conflict is recorded instead of one value silently replacing the other.

4. 4
   Policies are evaluated

   The proposed action is checked against gate policies derived from your controls: required state, open conflicts, thresholds and approvals.

5. 5
   A decision is made before the action

   The gate returns allow, warn or block. The decision does not depend on a model judgement, so it can be replayed.

6. 6
   A receipt becomes evidence

   Every decision is written to the append-only trail with what was asked, what was known and which rule applied.

See a decision being made

{
  "verdict": "BLOCK",
  "action": "release_payment",
  "decision_id": "gd_7Q2K9F",
  "reasons": [
    "Unresolved conflict on claim.amount ($48,200 vs $41,750)"
  ]
}