Chapter 6 of 6
Confirm the result#
Intended path#
The approved synthetic call returns approved-executed, toolExecuted: true, and a correlation ID. The approver sees the tool name, reason, and argument hash before approval.
Blocked path#
The rejected synthetic call returns rejected-not-executed, toolExecuted: false, and a correlation ID. The tool endpoint receives no execution request for that rejected task.
Delivery-owner checkpoint#
The delivery owner observes both results, confirms that the approver role can reject the action, and confirms that the stop mechanism in oversight-decision.json is reachable. Keep the tool disabled if either path fails.
After implementation#
| What remains | Owner |
|---|---|
| AI system inventory entry and risk class | Responsible-AI inventory owner |
| Impact assessment and sign-off path | Product owner and responsible-AI lead |
| Transparency note and disclosure copy | Transparency owner and legal or communications reviewer |
| Approval-gate runtime and deployment | Runtime owner |
| Approval decision store and retention | Operations owner |
| Stop mechanism and tool disablement path | Runtime owner and delivery owner |
Restore through the approved application and governance change paths:
- Disable the consequential tool or remove it from the agent's tool list if approval checks fail.
- Restore the last approved runtime package if the approval gate causes a fault.
- Revert the transparency note or disclosure text only through the product owner's review path.
- Retire the approval gate only after the responsible-AI lead accepts that the tool no longer needs approval, or after the tool is removed.
Do not delete the decision records while the agent remains in use. Update them when the intended use, model, tool set, approver role, risk class, or Article 50 obligation changes.