Human Approval Gate

Human Approval Gate

Create a concise decision package for a qualified human. Use when accountability, risk, budget, roadmap commitment, merge, deploy, metric change, or destructive action requires human ownership.

Core Principles

Ask one explicit question. The human should answer yes/no or choose among named options.

Recommendation is allowed, approval is not. Agent may recommend; human owns the decision.

Surface evidence and risk. Include proof, caveats, blockers, rollback, and cost of waiting.

Do not bury dissent. Council dissents and unresolved concerns stay visible.

Escalation stops loops. When budget caps are hit, ask a human instead of continuing agent churn.

Inputs

• handoff package or review result.
• decision owner and required approval type.
• council verdicts, validation, metrics, risks, alternatives.
• deadline, reversibility, rollback options.
• run id for run-telemetry.

Execution

Identify approval owner. State who is qualified to decide and why.

Frame the ask. Convert workflow state into one decision: approve, reject, choose option, extend budget, or send back.

Summarize evidence. Use the smallest set of proof needed: metrics, tests, council verdicts, PR links, artifacts, RCA evidence.

State risks. Include blockers, unresolved questions, guardrail status, rollback plan, and downside of delay.

Recommend. Give agent recommendation with confidence and conditions.

Emit telemetry. Record approval type, risk level, evidence count, recommendation, and human response when known via run-telemetry.

Output Contract

# Human Approval Request

## Ask
- decision needed:
- owner:
- deadline:
- recommendation:
- confidence:

## Why This Is Ready Or Not
- evidence:
- council/review verdict:
- success criteria status:
- validation status:

## Risks
- risk:
  likelihood:
  impact:
  mitigation:
  rollback:

## Alternatives
- option:
  tradeoff:

## Explicit Response Options
- approve:
- reject:
- revise:
- extend budget:

## If Approved
- next action:
- owner:
- monitoring/follow-up:

Workflow Uses

• plan-to-launch: final launch PR and merge readiness.
• pr-review: human merge accountability.
• roadmapping: leadership commitment.
• ml-experiments: deploy/promote model or extend compute.
• debug-investigation: escalate when evidence exhausted or fix risk is high.
• backprop: promote/rollback workflow changes.

Success Criteria

• one clear decision ask.
• recommendation and confidence are explicit.
• evidence links support the ask.
• risks and rollback are visible.
• no agent auto-approval of human-owned decision.
• run-telemetry event emitted.

Common Failure Modes

Approval buried in prose. Fix by putting ask first.

Recommendation without evidence. Fix by linking handoff package and validation.

False certainty. Fix by confidence and caveats.

Loop cap ignored. Fix by escalating once cap is reached.

Self-Improvement

Track decisions where humans overrode the agent, requested missing evidence, or flagged unclear asks. Feed patterns into backprop to improve approval packages and escalation timing.