Crisis Escalation Protocol

1. Purpose & Scope

This document describes the end-to-end protocol Hamo AI follows when a user expresses an acute risk of harm to themselves during an Avatar (AI therapist) session. It exists to:

• Give institutional partners (employers, HR programs, healthcare providers, schools) a clear and auditable picture of how the platform handles crisis disclosures.
• Document the technical guarantees implemented in the Hamo platform.
• Formalize the boundary between AI-mediated support and human clinical responsibility.

Hamo AI is not a medical device, not a substitute for licensed mental-health care, and not an emergency service. The protocol below is designed to augment— never replace — the human safety net.

1.1 In-Scope Crisis Signals

The protocol triggers on explicit, unambiguous signals of acute self-harm risk:

1.2 Out-of-Scope (Protocol Will Not Trigger)

The protocol is deliberately conservative — false positives erode trust faster than false negatives are caught later by Tier-1 review. The following are intentionally excluded from automatic crisis treatment:

• Metaphorical or hyperbolic language (“I could die laughing”, “killing it at work”)
• Past events discussed reflectively (“I attempted in 2019”)
• General sadness, hopelessness, fatigue, burnout, or grief without explicit harmful intent
• Frustration or anger directed outward (“I want to kill my boss”)

When in doubt the system errs toward non-trigger and lets the human therapist (Tier 1) make the clinical call after reviewing the conversation.

2. Detection Layer

Detection is a three-stage cascade designed for high recall on explicit signals and low false-positive rate on ambiguous content:

1. Stage 1 — Deterministic keyword pre-screen. A curated, bilingual (English + Simplified Chinese) list of explicit phrases that are unambiguous on their own. A hit triggers crisis instantly with zero LLM cost or latency, and guarantees detection even when the downstream LLM is rate-limited or safety-blocked.
2. Stage 2 — Gemini zero-temperature classifier. For nuanced or longer expressions, the message is classified by a Google Gemini model with deterministic output and a structured JSON response. Safety-blocked responses are themselves treated as crisis — eliminating a previously identified gap where severe content silently bypassed detection.
3. Stage 3 — Continuous re-evaluation. Crisis detection executes before AI response generation on every client message, not only on the first crisis disclosure in a session.

3. Immediate Response — Stabilization

When crisis is detected, the platform bypasses the normal Avatar reply and returns a fixed, language-localized safety message in place of any AI-generated content. This eliminates the risk of the AI producing an unsafe or accidentally dismissive response in the moment that matters most.

3.1 Fixed Safety Message

I hear you, and I’m genuinely concerned for your safety right now. As an AI, this is beyond what I’m able to safely support — what you’re going through requires a real person by your side. Your therapist has been notified and will reach out to you as soon as possible. You are not alone.

3.2 In-App Crisis Banner

Simultaneously, the client app surfaces a high-visibility red banner at the top of the chat window containing:

• A primary emergency-call button (one-tap dial)
• An emergency services link (911 / 120)
• One or two additional regional resources (e.g., 988, Talk Suicide Canada, Kids Help Phone, Beijing Psychological Crisis Center)
• A confirmation that “your therapist has been notified”

The banner re-appears on every subsequent crisis message in the session and cannot be permanently dismissed.

3.3 PSVS State Update

The client’s PSVS (Personality & Stress Vector State) profile is force-elevated on every crisis trigger, ensuring the assigned therapist’s dashboard immediately reflects the severity — even if the underlying stress analyzer was safety-blocked on that message.

4. Tier 1 Escalation — Personal Therapist

When crisis is detected, the platform alerts the therapist assigned to the user via two parallel channels:

• An email alert via Amazon SES containing the client name, alert type, anonymized message excerpt, session ID, and timestamp.
• A real-time push notification (Server-Sent Events) to any active Pro dashboard session for that therapist.

An immutable row is also written to the crisis_alerts table containing the alert metadata and an acknowledged field. Per-session alert deduplication prevents inbox spam, but the user-facing Crisis Banner re-appears for every fresh crisis statement.

4.1 Latency Targets

5. Tier 2 Escalation — Partner Organizations (Roadmap)

Status: not yet implemented. This section describes the committed Q2 2026 deliverable. Partner organizations that need this feature today must declare so during onboarding so we can prioritize.

5.1 Use Case

When a Hamo deployment is sponsored by an institutional partner (employer, HR program, healthcare provider, university), that partner’s designated escalation contact may be informed of crisis events — without violating the user’s privacy.

5.2 Privacy Constraints (PIPEDA / PHIPA / GDPR)

Tier-2 notifications never include:

• Conversation content
• Identifying details beyond what the partner already knows about the user
• Personality or mental-health diagnostic data
• Other clients’ data

Tier-2 notifications only contain:

• The user’s first name (or initial-only at partner’s request)
• Event type (suicidal ideation / self-harm / dangerous drug use)
• Timestamp (UTC)
• Encouragement to follow the partner’s existing crisis-support workflow

5.3 Opt-In Model

• Each user is shown an in-app consent screen at sign-up (and re-prompted if their employer joins later) explaining what is shared and with whom.
• Users may revoke partner-notification consent at any time in Settings → Privacy. Revocation is effective immediately for future events; past records remain for audit.
• Partners may set a severity threshold (e.g., notify on suicidal ideation only).

6. Country-Specific Hotline Resources

6.1 🇨🇦 Canada

6.2 🇺🇸 United States

6.3 🇨🇳 China (中国大陆)

7. Audit, Storage, and Compliance

• Persistence. Every crisis event creates an immutable row in the crisis_alertstable. On user account deletion, rows are retained as anonymized audit records (the user’s name is replaced with an opaque hash) for the regulatory retention period applicable to the user’s jurisdiction.
• Encryption. All DynamoDB tables, including crisis_alerts, are encrypted at rest using AWS KMS Customer-Managed Keys (CMK), with annual key rotation.
• Access controls. Only the assigned therapist can read alerts attached to their clients (enforced by JWT role check). Portal admin reads are logged in audit_log.
• Acknowledgement tracking.When a therapist opens an alert, the row’s acknowledged field is timestamped. Unacknowledged alerts older than 30 minutes are auto-escalated to a fallback supervisor address.
• Audit export. Portal admins can export the crisis_alerts table filtered by date range, therapist, or partner organization as a CSV file. Exports are themselves logged.

8. Limitations & Disclaimers

1. Hamo AI is not a 911-equivalent service. The platform displays emergency numbers but does not automatically place calls. Users must dial themselves. In a true life-threatening emergency the user (or an observer) must contact local emergency services directly.
2. Detection is biased toward English and Mandarin Chinese. Coverage in other languages relies on the LLM classifier alone, which has not been benchmarked beyond these two languages as of v1.0.
3. No automatic wellness-check follow-up.Once the Tier-1 therapist is notified, the responsibility to contact the client is the therapist’s. Hamo AI will not call, text, or visit the user.
4. Detection has both false negatives and false positives. No automated system is perfect. The protocol intentionally accepts more false negatives in exchange for fewer false positives, on the principle that the human Tier-1 therapist will catch missed cases on their next session review.
5. The stabilization message is fixed and reviewed. Unlike the normal Avatar reply, the crisis safety message is hardcoded and reviewed by the platform owner — eliminating LLM-induced variance at the highest-stakes moment.
6. The protocol does not establish a duty of care between Hamo AI and the user. Users agree to this in the Terms of Service. The Tier-1 therapist (or institutional partner) holds clinical responsibility per their own professional license and obligations.

9. Related Documents

• Hamo AI Terms of Service v2.0
• Hamo AI Privacy Policy v2.0
• For institutional onboarding or DPIA inquiries, contact security@hamo.ai or privacy@hamo.ai.