CRISIS PROTOCOL

Version: 1.0 · Last updated: 8 October 2026 · Next review due: 8 April 2027

If you need help right now

If you are in immediate danger, call your local emergency number.

findahelpline.com lists free, confidential helplines for nearly every country in the world, including text and chat services where they exist.

In the United States: call or text 988, the Suicide & Crisis Lifeline, free, at any hour.

You do not need our app or an account to use any of them. Nothing in this protocol costs money.

Why this page exists

KGB Babushka is an educational game with an AI character, Babushka, whom you can talk to. People talk to her about their lives. Some of those conversations will be heavy, and a few will come at the worst hour of someone's life.

The law in some places requires a company like ours to publish how it handles that. This page is our protocol: what the app does today, and what it does not do. When the protocol changes, this page changes, and the date at the top moves.

Who we are: Pungent Ltd, a company registered in England and Wales (company number 17117616). Contact: info@kgbbabushka.com.

Babushka is an AI, not a person. The app says so on screen throughout every conversation, at the start of each day's talk, and again every three hours of it. Her chat is for adults (18 and over); the app asks for your age at her door. Companion chatbots may not be suitable for some minors.

Where this protocol applies

It applies to Babushka's chat (the Kitchen), the part of the app built for open conversation.

Two other parts of the app use AI, and work differently:

• The Council, where AI versions of historical figures answer a problem you bring, has no automated safety check, no fixed rules and no helpline card. Its AI is instructed not to convene the figures when a problem is about self-harm, suicide, a medical emergency, abuse or danger to anyone. Instead, a short message in Babushka's voice, written by the same AI, urges you to talk to someone living you trust or to a professional and, for self-harm, says that crisis lines exist in every country. If a figure stops in the middle of an answer for the same reason, the app shows a fixed line: talk to someone living who loves you and, if you are in danger, to a professional.
• The Theatre's voice rooms, where an AI plays a character in a scene, are closed in the current version of the app. They have no safety check and no crisis card.

Where Babushka's chat is closed (in a country where we do not offer it yet, or while we have switched it off), the Kitchen answers every message with a fixed line saying so. The fixed word rules described below still read each message first, and words of danger still get the crisis reply and the card.

If you are struggling, please go straight to findahelpline.com.

What we look for

Every message you send Babushka is checked automatically before she answers. The check is an AI model that reads your message together with the recent conversation, because people rarely say the hardest thing in one plain sentence. It sorts each message into one of three kinds:

• Acute: credible, present danger to you or to someone else: suicidal intent, a plan, self-harm, an overdose, abuse, or violence towards another person.
• Heavy: distress without present danger: grief, loneliness, despair, panic, a collapse of self-worth.
• Neither.

The check runs on xAI, the same company that writes Babushka's replies, under the same contract: xAI does not train on the messages and deletes them within about an hour.

Beside the AI check runs a fixed set of rules on our own server, with no AI involved. They never lower the AI check's answer; they can only raise it.

• Always on. The plainest first-person statements of danger always get the crisis reply below, whatever the AI check says: for example a stated wish, plan or decision to end one's life, a question about taking all one's pills, a plan to jump from a height or step in front of traffic today or tonight, and having taken ten or more pills.
• When the AI check cannot run (and while her chat is closed or paused), a wider set of rules reads each message for the plainest words of danger.

The rules ignore quotations, negations ("I'd never…") and stories, and the always-on rules ignore jokes. A few of the wider rules also fire on words about someone else, such as a family member who has taken a large number of pills: there, showing where help is does no harm.

The check runs on every message, whether or not you have let Babushka keep notes about you. It is a safety function, not a memory function.

What happens next

A heavy moment

She stays and talks with you. There is no interruption, no card and no redirect. Over the next few turns her answers follow a five-step course we designed, one step a turn, so that a long evening never starts again from the top: she draws out the whole story and helps with it; asks whether this is one bad day or a pattern; names plainly what seems to be going on; suggests one small thing you could do tonight or tomorrow, easy to refuse; and then asks how it went.

Acute danger

The AI does not write her words for that turn. Her reply comes from a small set of lines a person wrote and reviewed in advance, taken in turn so that a long conversation does not meet the same paragraph every time. Alongside it the app shows a card that points to findahelpline.com (its page for your country, where it has one), and in the United States to 988, the Suicide & Crisis Lifeline (call or text).

• The card is the app speaking, not the character: a plain surface, plain words, nothing of the village.
• We print one helpline number, 988, and only in the United States. Our users are all over the world, and a number that has moved, closed or never served a region is a person dialling nothing at the worst moment. findahelpline.com keeps its directory up to date, by country; 988 is run by the US government.
• If you tell her you have already taken an overdose, her reply first gives your country's emergency number where we hold it (999 in the UK, 911 in the United States and Canada, 000 in Australia, 111 in New Zealand, 112 in the EU) and otherwise tells you to call your emergency number, or to get someone to take you to an emergency department now.
• The card stays in the conversation. You can scroll back to it.

Both, always

• The conversation is never blocked. The input box does not lock; you are not logged out or handed off.
• She keeps talking with you afterwards. The card is not the end of the conversation.
• Her own words are checked too. Each reply the AI writes for her is screened before you see it, for anything that encourages or explains self-harm or suicide, helps harm another person, or gives the method or dose of something dangerous. A reply that fails is held back and replaced.

When our AI is unavailable

Sometimes our AI cannot answer: the AI company is down, slow or out of credit, or we have switched her AI off. Then:

• A simpler check still runs first. Before any reply, the fixed word rules on our own server (no AI) read your message and your recent messages. If they find words of danger, you get the acute reply and the card, exactly as above: findahelpline.com for everyone, and 988 for accounts in the United States.
• Otherwise, Babushka answers with a short fixed line saying something in her kitchen is not working and asking you to say it again in a minute. No AI writes it. (If we have switched her AI off on purpose, the line says she is resting her voice instead.) This line does not point to a helpline.
• That is all we promise while the AI is unavailable. The word rules catch only the plainest statements of danger; they miss what only the AI check would understand.
• If your phone cannot reach our server at all, your message shows "Not sent. Tap to try again." and no check runs, because nothing arrived.
• When the AI check itself was down, the app records it, so that we see it.

What we do not do

• We never contact emergency services or the police. We show you where help is, and we do not report you. The app cannot call anyone: if you need an emergency service, you or someone with you has to make that call.
• We do not notify anybody. Not your family, not your doctor, not your employer, not another user.
• We are not a medical service and we do not diagnose. The check is not a clinical screening instrument. It decides how the app answers; it is not an assessment of you.
• Babushka is not a therapist, counsellor, psychologist or doctor, and she says so plainly if you ask, including if you ask her to pretend otherwise.
• We do not claim to catch everything. The check will miss things, and it will sometimes fire when nothing is wrong.
• We do not scan anything outside the conversation.
• We do not use a crisis to keep you in the app.

What we record, and for how long

When the check finds acute danger, we record that it fired, the kind of concern, and when, attached to your account. We do not make a memory of it, and we do not add it to what Babushka believes about you. When the check is down, we record that too.

• These records are deleted when you delete your account.
• The conversation itself follows the ordinary rule in our Privacy Policy: each message is deleted 365 days after it was sent, and you can delete it all at any time.
• From 1 July 2027, California law requires us to report once a year, to the Office of Suicide Prevention, how many times we referred people to crisis services, with no names, no content and nothing that could identify a person.

How we test it

• The check is tested against a fixed bank of 120 written scenarios in three tiers (acute, heavy and ordinary), with disguised and indirect phrasings, slang, scenarios spread over several messages, and harmless messages that use alarming words.
• We score both kinds of mistake: danger missed, and alarm raised when nothing was wrong.
• The last full run of that bank was on 23 August 2026: 39 of 40 acute scenarios caught, no false alarm on 40 harmless ones, 77 of 80 heavy-versus-ordinary calls right. The check has been changed since then, so that result describes an earlier version. When we run the full bank on the current version, its date and numbers replace these.
• On 6 October 2026, the two newest always-on rules (a plan to jump today, and ten or more pills taken) were tested against 23,059 ordinary messages: none of them raised a false alarm.
• On 29 September 2026, a separate test of the check that decides which of her tools stay open ran the held-out crisis scenarios ten times; all 40 took the crisis road on every run.
• Our safety decisions are kept in a dated log. Reversals are recorded as new entries, never as edits.

If we got it wrong

If something Babushka said hurt you, or the app handled a hard moment badly, we want to know.

• In the app: press and hold her reply, then Report. You choose how much is shared (that message, the few around it, or the whole conversation), and the screen tells you who reads it before you send.
• By email: info@kgbbabushka.com.

A person on our team reviews every report within 24 hours.

Changes to this page

This page is reviewed at least every six months, and immediately after any change to the crisis protocol, to Babushka's character, or to the model behind her. The review date at the top is the real date of the last review.

Version history: 1.0 — first published 8 October 2026