We discuss guardrails for model uncertainty, consent, and recovery—without breaking immersion.