Escalation to Human
Domain 6 is unscored but frequently tested. It governs when an agent should hand off to a human, what "wrong" escalation reasons look like, and what a correct handoff carries. The grounding comes from the Netlight study guide plus the Model Spec / Claude's Constitution (principal hierarchy Anthropic > operators > users, reversibility-weighted harm, least-drastic-action).
Concepts
- Escalation triggers — the 3 valid triggers: explicit human request, ambiguous/silent policy, no meaningful progress (complexity alone is NOT one).
- Wrong escalation signals — always-wrong bases: sentiment/frustration, confidence scores, complexity alone.
- Explicit-request-overrides rule — "I want a human" → escalate immediately, no investigation, overrides everything.
- Structured handoff summary — the 5 required fields (verified ID, root cause, attempts+outcome, amounts/orders, recommended action); never dump a transcript.
- Minimal footprint principle — prefer reversible, do less when scope unclear, abandon vs proceed recklessly.
- Human-in-the-loop calibration — auto-resolve >90% confidence with field-level scores; validate per segment.
- Limits of automated retry — 2-3 attempts then fail fast; retry is useless for missing information.
- The Escalation Handoff — the Playbook's decision-tree realization of these rules.
- Claude's Constitution