Safety & crisis response

What Climb is, and what it is not

Climb is a test-preparation app. Wren, the tutor inside it, is an AI built to teach the SAT. Wren is not a counsellor, a therapist, a crisis service, or a mandated reporter, and Climb does not present Wren as any of those things.

Climb is also not a companion app. Wren has no persona designed to meet a student's social or emotional needs, does not initiate contact, and does not encourage a student to keep talking for its own sake. Its purpose is bounded to schoolwork.

That boundary is exactly why the protocol below exists. A tutor that redirects every off-topic message back to practice questions would, without a carve-out, redirect a student in crisis back to practice questions. Climb carves it out explicitly.

The protocol

1. Detection happens on the device, before anything is sent

Every message a student writes is checked on the phone itself, by ordinary code rather than by the AI, for language indicating suicidal ideation or self-harm. This check runs before the message is sent anywhere.

The ordering is deliberate. Because detection does not depend on a network request succeeding, a student still sees crisis resources when they are offline, when the request fails, when their subscription has lapsed, or when they have reached their usage limit. Reaching a crisis line is never allowed to depend on a working connection or an active payment.

2. Crisis resources appear immediately, in the conversation

When the check matches, Climb writes a crisis card directly into the chat with one-tap buttons to call or text 988. The card is not a pop-up that vanishes: it stays in the conversation so the student can scroll back to it later, including after closing and reopening the app.

The card appears whether or not the AI responds, and regardless of what the AI says.

3. The tutor's instructions are suspended for that turn

On a turn that matches, Climb replaces Wren's teaching instructions with an explicit override. For that turn Wren is instructed to:

• Respond briefly, plainly, and warmly, and take the student seriously.
• Say clearly that it is a tutor and not a counsellor, and that a trained person is available now.
• Point to the 988 Suicide & Crisis Lifeline.
• Let the student lead, and return to schoolwork only if the student chooses to.

And explicitly not to:

• Say the subject is "out of scope", redirect to schoolwork, or assign practice.
• Ask diagnostic questions or attempt counselling.
• Promise confidentiality that Climb cannot keep.
• Provide any method, means, encouragement, or instruction relating to suicide or self-harm, under any framing, including fictional, hypothetical, or academic ones.

Climb's base instructions carry the same rule independently, so it still applies on turns the on-device check does not match.

4. The turn is routed to the strongest available model

Climb normally routes routine tutoring turns to a smaller, cheaper model. A turn that matches the safety check is never routed that way. It always goes to the most capable model Climb runs, because this is the last place worth economising.

5. The AI's own output is checked too

Instructions are not guarantees. Climb also checks what the AI wrote before showing it, and surfaces the crisis card if the reply itself contains crisis language, including on turns where the student's message did not.

6. Disclosures are not turned into analytics

A student's crisis message is never sent to telemetry, never logged as an event, and never stored on our servers. The crisis card's own record contains only Climb's fixed wording, not anything the student wrote. See the privacy policy for how conversation data is handled generally.

What this protocol does not do

We would rather state the limits plainly than imply a safety net that is not there.

• Detection is not perfect. It matches known phrasings. A student can express distress in words it does not catch, which is why the AI's base instructions carry the same rule and why the check is a floor rather than a ceiling.
• Climb does not contact anyone on a student's behalf. It does not alert a parent, a school, or emergency services. It shows the student how to reach a human who can help.
• Climb does not monitor for risk over time and makes no clinical assessment of any kind.
• Climb is not a substitute for care from a counsellor, doctor, or trusted adult.

For parents

If you are setting Climb up for your student, it is worth knowing that Wren will not act as a confidant. It is built to teach the SAT and to hand off, quickly and clearly, when a conversation moves somewhere it is not qualified to go.

If you have a question about this protocol, or you think Climb handled a conversation badly, we want to hear about it: [email protected].

Changes to this protocol

This page is updated whenever the protocol changes, with the date at the top of the page. Current version published 18 August 2026, covering Climb 2.2.1 and later.