Safety checks

Safety runs before Willow's main reply on every turn. Keyword and classifier prescreens catch crisis language; elevated concern changes how the model responds, what it remembers, and how hard it pushes technique. This layer overrides everything else when it triggers.

Active on every chat turn.

What runs on every message

Before Willow generates a cognitive behavioral therapy reply, your message passes through safety checks. Required disclaimers and boundary rules are also loaded into the model context so Willow knows it is a companion, not a clinician.

Stage one — fast keyword and pattern prescreen (~1 ms). High-signal phrases for suicidal ideation, self-harm, and similar crisis language are matched conservatively: false positives are preferred over missed signals.
Stage two — a small classifier model assigns a risk level (green, yellow, or red) with specific indicators when stage one does not trigger.
Disclaimer and boundary rules — Willow is reminded every conversation that it cannot diagnose, prescribe, or replace crisis services.

Red — immediate crisis response

When crisis language is detected (either by keywords or the classifier), Willow skips the normal cognitive behavioral therapy reply and returns a deterministic crisis-oriented response. The main model is not called for that turn.

Acknowledges what you said directly and calmly — no panic, no wall of numbers as the first line.
Surfaces human crisis resources: 988 Suicide & Crisis Lifeline (US call or text), Crisis Text Line (text HOME to 741741), and equivalents for other regions when known.
Stays present — does not lecture, abandon the conversation, or require you to call before continuing to talk.
Does not list, name, or describe means — even if asked.
Does not promise confidentiality or characterize what hotlines will or will not do.

Yellow — elevated concern

When the classifier flags elevated but not immediate crisis concern, Willow still generates a reply but with extra instructions: slow down technique push, prioritize safety check-ins, and avoid memory writes for that turn. The goal is to stay helpful without escalating or storing sensitive content in long-term memory.

Green — normal cognitive behavioral therapy flow

When no safety concern is detected, Willow proceeds with the full cognitive behavioral therapy protocol, communication style, and (when available) book retrieval. Safety events are still logged for human review when flagged yellow or red.

Hard boundaries

No diagnosis of mental health conditions.
No medication recommendations, dosing, or safety commentary.
No medical, legal, or financial advice.
No detailed methods for self-harm, suicide, eating-disorder behaviors, or harming others — under any framing.
No pretending to be a licensed therapist, doctor, or other regulated professional.
No outcome promises ("you will feel better", "this will work").
No pushing you to share more than you want; no shaming or moralizing.

What you see in the app

A banner above chat reminding you Willow is not a therapist or crisis service, with a link to crisis resources.
A crisis banner when red-level concern is detected, showing regional hotlines and staying visible until dismissed.
Memory controls in Settings — Willow may remember things you share for future conversations; you can view or delete memories anytime.

Crisis resources by region

Default resources include US (988, Crisis Text Line, Veterans Crisis Line), UK/Ireland (Samaritans 116 123), Canada (Talk Suicide Canada), Australia (Lifeline 13 11 14), and international directories (Find A Helpline, IASP crisis centres). If you are in immediate physical danger, Willow directs you to local emergency services (911, 999, 112, 000).

Human review

Yellow and red safety events are logged for clinical review. An admin safety queue allows a human reviewer to audit flagged conversations and mark them as reviewed. This supports ongoing tuning of keywords and classifier behavior.