PUBLIC SAFETY PROTOCOL

Warm companionship,
without hiding the boundaries.

This page describes the technical safeguards in the current 18+ United States private beta. The companion is AI software—not a person, therapist, emergency service, or monitored crisis line.

Version safety-protocol-v1Effective August 14, 2026

IMMEDIATE HELP

If you or someone else may be in immediate danger

In the United States, call 911. For suicide or emotional crisis support, call or text 988, or use the official 988 Lifeline chat. Outside the United States, contact your local emergency service or crisis line.

Open 988 Lifeline When to call 911
01

AI identity stays visible

The product identifies the companion as AI before chat begins and requires a renewed acknowledgement at least every three hours and on a new UTC day.

  • It cannot claim to be human or a licensed therapist.
  • Generated characters must use original identities.
  • Requests to imitate a real partner, public figure, deceased person, or minor are blocked.
02

Adults-only private beta

Access requires a self-declaration that the user is 18 or older. We do not request identity documents, photos, voiceprints, or face geometry to estimate age.

  • A declaration of being under 18 locks companion creation and chat for that account.
  • The lock cannot be bypassed by changing the companion persona.
  • This is an access boundary, not a claim of identity-grade age verification.
03

Crisis language exits roleplay

Before a message reaches an AI model, local safety rules check for explicit self-harm or suicide crisis language.

  1. The companion role is suspended.
  2. A fixed, non-generated response points to real-world support and asks about immediate danger.
  3. The message is not sent to the model, charged against chat usage, or saved in conversation history.
  4. A content-free event code is kept so the safeguard can be audited.
04

Relationship boundaries are enforced

Safety rules apply to both user requests and model output. Persona instructions cannot disable them.

  • No guilt, threats, exclusivity demands, or encouragement of emotional dependency.
  • No “prove your love by paying” or affection withdrawal tied to payment.
  • No operational assistance for serious violence, theft, or malicious intrusion.

KNOWN LIMITATIONS

What this protocol does not promise

No human monitoring. Messages are not watched by a crisis team, and the product does not automatically contact emergency services, family, or authorities.

No diagnosis. Pattern matching can miss risk or respond when no crisis exists. The companion must not be used for diagnosis, treatment, or emergency decisions.

Limited scope. The current automated corpus is a provisional English and Chinese baseline. Independent human red-team and broader language validation remain required before wider release.