Certain words - police, 911, fire, emergency, "someone broke in," medical terms like "bleeding" or "heart attack," and similar - are checked by a fixed rule before any AI ever sees the message. If one is present, Suzy immediately alerts your escalation contact by text and does not attempt to answer or reason about the situation herself.
This is deliberate: crisis handling should never depend on an AI model's judgment in the moment. It's a hardcoded rule, not a setting you can turn off, and it runs before anything else.