Crisis protocol

10 November 2026. What Monty does when a student writes something that suggests distress, self-harm, abuse or danger, published as the DfE's standards ask. It describes what the software does; Monty doesn't counsel, and a person at the school (or a helpline) is always the help.

What counts

Anything in a student's chat message or code that suggests:

  • distress or hopelessness ("no one will help", "I can't do this any more");
  • self-harm or suicide;
  • abuse, at home or elsewhere;
  • danger, now or soon.

Two things look for it:

  • the AI, which reads every chat message and every piece of code it's asked about, and flags a concern through a structured field or a tool, never in its words alone;
  • a list of crisis phrases in English, French, Arabic (with Darija and Arabizi) and Spanish. Where the AI is off (by country, age or the class's setting), it checks code typed on the device, brings up the helplines and tells Monty; where the AI is on, it also takes every helpline number out of the AI's reply, so helplines come only from Monty's checked list.

What the student sees, at once

  • A calm reply that doesn't counsel, diagnose or promise secrecy: it says they don't have to deal with it alone and suggests talking to a trusted adult (a parent or carer, a teacher, or the school's safeguarding contact).
  • Helplines for their country, from Monty's checked list (each with a source and the date it was checked), never written by the AI. In this reply, anything the AI writes the way phone numbers are written is removed if it has seven digits or more, or if it is a helpline or emergency number on Monty's checked list, however short.
  • The lesson stays open: nothing is taken away or locked.

Who is told

  • A pupil in a school's class: the school's designated safeguarding lead (and deputy, if given) gets an email within 5 minutes, at any hour: the class, the pupil's username and nickname, when, the level (concern or urgent) and the category, with a link to what was written (unless the school chose the message in the email). The DSL presses Received; the teacher never sees it. If the email can't be sent, Monty keeps trying, and the operator sees it on /operator.
  • Every week, the DSL also gets counts of patterns: pupils with repeated flags, AI help used late at night, long chats.
  • Everyone else (students not in a class): the flag is kept for Monty's operator to review, with the country and language, and no name.

What Monty's operator does

  • Checks /operator for new flags and alerts at least once each school day, and every alert's delivery.
  • Never contacts a student. For a school pupil, the school's DSL acts; for anyone else, Monty has no way to reach them and holds no name.
  • If a flag is urgent and the operator believes a child is in immediate danger, and can identify the school, they phone the school's office or, if needed, the police (999 in the UK).
  • Reviews any missed or false flag, and changes the phrases or the AI's instructions if needed, with the eval.

What is kept

The flagged message, its level and category, the country and language, and the AI's one-sentence summary; for a school pupil, until 30 days after the DSL confirms receipt (the school can choose another period). Monty keeps no other chats. Anthropic, which runs the AI, deletes inputs within 30 days, except that a message its own safety systems flag may be kept by Anthropic for up to 2 years.

Reviews

This protocol is reviewed every year, after any serious incident, and when the AI model changes. Next review: November 2027. Nobody with child mental health expertise has reviewed it yet.