Human Oversight

The human-oversight policy is a focused output-phase escalation control. When action: escalate is configured, the gateway marks the result as ESCALATE, returns a successful response with null assistant content, and 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
require_human_for	string[]	Accepted targeting metadata; the current evaluator does not inspect it.	[]
action	escalate | block	Only an explicit escalate is effective today. block validates but behaves like no policy action.	no runtime action when omitted
confidence_threshold	number	Accepted range is 0–1; the current evaluator does not score confidence.	0.5
default_assignee	string	Accepted assignment metadata; routing currently comes from provider escalation routing instead.	—
timeout_seconds	integer	Accepted range is 1–604800; the current evaluator does not wait for review.	3600

How it works

1. Output phase only: human-oversight runs after the model response is available.
2. Escalation trigger: the gateway acts only when policy.human-oversight.action equals escalate.
3. Response shape: the gateway returns a successful chat-completion style body with choices[0].message.content = null and finish_reason = "content_filter".
4. Event metadata: the gateway emits a decision event with policy_kind = "human-oversight" and reason_code = "oversight.required".
5. Routing hint source: if the requested model resolves to escalation routing in the provider configuration, the gateway attaches team_id and user_id hints to the event metadata. Those routing hints do not come from the policy block itself.

Configuration boundaries

Use action: escalate when every response reaching this chain entry should stop normal assistant delivery and hand the request to a review workflow. The other accepted fields are not runtime targeting, assignment, or timeout controls yet; setting them does not narrow when escalation happens.

If you need reviewer assignment or team routing, configure it through provider-level escalation routing or the downstream review workflow rather than the human-oversight policy block.

Valid example

pack:
  name: eu-ai-output-gate
  version: "1.0.0"
  enabled: true

policies:
  chain:
    - quality-scorer
    - human-oversight

policy:
  human-oversight:
    action: escalate

Best practices

• Treat human-oversight as a simple escalation switch, not a classifier or queue-definition system.
• Pair it with quality-scorer or citation-verifier when you need a deterministic reason for why an output was escalated.
• Add audit-logger only when the decision stream should mark audit logging as an active chain control; that policy does not persist evidence itself.
• Verify the downstream system consuming decision events understands the ESCALATE verdict.

Next steps

• Quality Scorer — Gate outputs before escalation
• Citation Verifier — escalate when groundedness is not strong enough
• Flagged Review Configuration — route borderline output through a reviewer
• Audit Logger — mark audit logging as active in the decision stream
• Trail and Audit Evidence — inspect and verify persisted evidence