Human Oversight

The human-oversight policy is an output control. With action: escalate, the gateway marks the result as ESCALATE. It returns a response that succeeds with null assistant content. It also records the escalation in the decision event.

Configuration

pack:
  name: human-oversight
  version: "1.0.0"
  enabled: true

policies:
  chain:
    - human-oversight

policy:
  human-oversight:
    action: escalate

Fields

Field	Type	Description	Default
action	escalate	The only supported runtime behavior. It escalates each response that this policy evaluates.	no runtime action when not specified

How it works

1. Output phase: human-oversight runs after the model response is available.
2. Escalation trigger: The gateway acts only when policy.human-oversight.action is escalate.
3. Response shape: The gateway returns a chat-completion body that shows success.
4. This body has choices[0].message.content = null and finish_reason = "content_filter".
5. Event metadata: The gateway emits an event with policy_kind = "human-oversight" and reason_code = "oversight.required".
6. Routing hints: Provider escalation routing can add team_id and user_id hints to event metadata.
7. These hints do not come from the policy block.

Configuration boundaries

Use action: escalate when each response must stop usual delivery and go to a review workflow.

Use stream: false. The active SSE output path does not run this escalation evaluator.

For reviewer assignment, category targeting, or team routing, use provider escalation routing. Alternatively, use the review workflow after the gateway. Do not configure these functions in the human-oversight policy block.

Correct example

pack:
  name: manual-review-route
  version: "1.0.0"
  enabled: true

policies:
  chain:
    - human-oversight

policy:
  human-oversight:
    action: escalate

Best practices

• Use human-oversight as an escalation switch. It is not a classifier or queue system.
• Put it only on a route or conditional chain where each response must have review.
• An earlier block or escalation stops the stage. A subsequent human-oversight entry does not run in that case.
• Add audit-logger only when the decision stream must show audit logging as an active chain control. That policy does not store evidence.
• Verify the downstream system consuming decision events understands the ESCALATE verdict.

Next steps

• Escalation Routing Configuration — add team or user routing hints
• Flagged Review Configuration — route borderline output through a reviewer
• Audit Logger — mark audit logging as active in the decision stream
• Trail and Audit Evidence — examine and verify recorded evidence