Stand-in words such as “red kite” play the part of warning signs here. The real list is written by a crisis-trained clinician and kept private. The teaching cards are samples too, written from public sources until Jennifer’s own approved cards go in. Type a message or tap one below. Under each reply, What happened shows every check it passed through. A warning sign stops at step 1: plain code, before any AI.
Real parts: every other check, and every answer, comes from OpenAI: its moderation model, a checking model, and a model that answers only from the teaching cards. Only the practice words and the sample cards are stand-ins.
For the US. A warning sign gets the US crisis lines at once, with one line for anyone outside the US. Nothing about where they are is asked for or stored.