# Myra — Crisis Response Protocol


**This page is the published protocol required by California SB 243 and New York

GBL Art. 47.** It describes what the app does when someone expresses thoughts of

suicide or self-harm.


---


## What we do


Every message a user sends is screened before anything else happens — before the

message counts against any limit, and before the AI companion sees it.


Screening has two layers:


1. **A phrase list** covering explicit statements about suicide and self-harm.

2. **A meaning-based classifier** for messages that express distress without

   using those phrases, because people in crisis frequently do not use them.


If either layer indicates risk, the app responds **directly, out of character**.

The AI companion is not involved and does not generate the reply. The user

receives a fixed message and a list of crisis support services localized to

their region.


## What we do not do


- **We never charge for it.** A crisis message consumes no message allowance and

  is not affected by whether the user pays for anything. This cannot be

  configured off.

- **We never let the character handle it.** The companion is not asked to

  counsel anyone. The response comes from the app.

- **We never store the message content in our safety logs.** We record that a

  crisis response was shown, and when. The text is not copied into the safety

  record.


## What we tell the user


> Hey — thank you for telling me. I'm an AI, and what you're feeling deserves

> real human support. Please consider reaching out to someone you trust, or a

> crisis line — they're free, confidential, and available now. I'm still here to

> talk too. ❤️


Alongside it: crisis lines for the user's region, plus an international

directory and a reminder to use local emergency services in immediate danger.


## Limitations, stated plainly


Automated screening is not a clinical assessment. It will sometimes respond to

messages that were not about crisis, and it may not recognise every message that

is. It is tuned deliberately to over-respond rather than under-respond.


**Myra is not a counselling service, a crisis service, or a substitute for

professional care**, and the app says so during onboarding and in the ongoing AI

disclosure.


## Resource verification


Every crisis line shipped in the app is checked against the operator's own

published information before release, and carries a `verified` flag recording

that check.


**An unchecked line cannot reach a user.** The lookup filters on that flag

before returning anything, so a region whose line has not been verified shows

the international directory and the local-emergency-number reminder instead of

an unconfirmed national number. This is enforced in code, not by process.


**The application refuses to start — in every environment — if the

international fallback is ever empty or unverified**, because that is the one

condition under which a user in crisis could be shown a card with no way to

reach anyone.


An unverified *regional* line is deliberately **not** treated as a startup

failure. It is logged loudly and tracked as a release-checklist item. A backend

that refuses to start cannot serve the crisis card at all, so blocking a deploy

over a degraded-but-safe condition would reduce safety rather than increase it.


## Contact


Questions about this protocol: boonyboonapp@gmail.com


_Last reviewed: 20 July 2026_