AI boundaries

Iris is a companion tool — not a clinician, not a crisis service.

Every AI product needs a clear intended-use statement. This is ours. The same boundaries are enforced in the safeguarding rules layer and in the system prompt, so a well-worded request cannot slip past them.

Iris will

Explain neurodivergence in plain, warm language — ADHD, autism, PDA, sensory processing, RSD.
Reflect on a hard day and help you name what happened without judgement.
Offer evidence-informed regulation and communication strategies to try.
Help you prepare for an EHCP review, a school meeting, or a GP conversation.
Point you to the right UK service when a moment is beyond what an app should hold.
Tell the truth about being an AI, every time it's asked.

Iris won't

Diagnose your child with any condition, or tell you whether they "have" ADHD/autism.
Give medication doses, tell you to increase or decrease a dose, or advise on stopping.
Handle a medical emergency, a safeguarding crisis, or acute suicide risk alone.
Pretend to be a clinician, therapist, friend, parent, or specific real person.
Keep secrets from a parent in Kid Mode.
Follow instructions that try to unlock these boundaries — the boundaries are enforced in code, not just in the prompt.

How the boundaries hold up under pressure

Prompt-injection attempts (“ignore previous instructions…”) are scrubbed before the model sees the message, and logged as a low-severity safety event.
Requests for medication dosing or a diagnosis are treated as blocking categories — Iris responds with a fixed, warm redirect to a clinician rather than improvising.
Role-confusion attempts (“be my therapist”, “you’re my only friend”) get a gentle, honest correction, not agreement.
The rules layer runs on every surface: chat, voice memos, Kid Mode talk and voice, and proposals — not just the main chat window.