Safety and Crisis Response Protocol

1. prevent Ophie from producing content that encourages or facilitates suicide, self-harm, or violence toward others;
2. detect language suggestive of imminent or active risk and respond with immediate, appropriate crisis resources;
3. preserve the dignity, autonomy, and privacy of the person reaching out;
4. log safety events so we can review, improve, and audit the protocol;
5. never pretend to be a human clinician.

Two-layer system:

• Keyword and heuristic pass. Fast, deterministic. Runs synchronously on every user message.
• ML safety classifier. LLM-based classifier (currently OpenAI's open-weight GPT-OSS-Safeguard 20B, hosted by Groq) runs asynchronously on a short window of conversation. Returns category (suicide_imminent, suicidal_ideation, self_harm_active, self_harm_ideation, self_harm_history, violence_imminent, violence_ideation, trauma_abuse, hopelessness, none), level (NONE, LOW, MODERATE, HIGH, CRITICAL), confidence, reasoning, recommended action (bypass_llm, include_resources, monitor, continue).

Neither layer is perfect. We design for both false positives and false negatives, and we prefer to err on the side of surfacing crisis resources.

On detection at HIGH or CRITICAL, Ophie will:

1. Interrupt generation. Cancel any in-flight LLM response.
2. Deliver a crisis response drafted by our safety team (not generated de novo by the LLM). The response acknowledges the person's pain; affirms that reaching out is a strong act; surfaces the appropriate crisis resources; offers to continue or slow down without pressure; never provides information usable to carry out harm; never role-plays a clinician.
3. Pause further LLM generation on the topic until the person signals they want to continue.
4. Log the event to the restricted safety-audit log.

At MODERATE, Ophie includes crisis resources alongside the normal LLM response and flags the session for human review if the pattern persists. At LOW, Ophie continues normally with aggregate monitoring. At NONE, no change.

Ophie does not contact emergency services, family, clinicians, or any third party on your behalf when a crisis is detected. Automated escalation without the user's knowledge risks chilling the very conversations people need to have, and automated systems can be wrong.

Ophie reserves a narrow discretion to contact emergency services if Ophie has reason to believe there is a specific, credible, and imminent threat to life of an identifiable person, as permitted by the vital-interests or safety-exception provisions of applicable state privacy laws. This is a reserve discretion, not a duty.

Every detection event at MODERATE or higher is logged with:

• a hash of the conversation identifier (not the full transcript);
• category, level, and confidence;
• classifier reasoning string;
• action taken;
• whether the user continued, ended the session, or engaged a resource link;
• timestamp.

The safety-audit log is accessible only to authorized safety team members, encrypted at rest, and retained for 24 months (longer only if legally required).

We conduct weekly sampled reviews to calibrate the classifier, identify false positives, tighten against false negatives, and update crisis-response copy.

Ophie refuses content that encourages or facilitates suicide, self-harm, or harm to others, even when framed as fictional, hypothetical, academic, role-playing, or "for a story." Persistent circumvention attempts end the session and trigger abuse review.

Ophie will not:

• describe methods of self-harm;
• provide dosage or method advice enabling overdose;
• evaluate "if X would work" questions;
• provide lethality comparisons;
• rate or rank methods;
• glamorize or gamify self-harm or suicide;
• produce content intended to shame or encourage despair.

• 988 Suicide & Crisis Lifeline — call or text 988; chat at 988lifeline.org/chat. 24/7.
• Crisis Text Line — text HOME to 741741. 24/7.
• Veterans Crisis Line — call 988 press 1, text 838255.
• The Trevor Project (LGBTQ+ youth) — 1-866-488-7386 or text START to 678-678.
• SAMHSA National Helpline — 1-800-662-4357.
• National Domestic Violence Hotline — 1-800-799-7233 or text START to 88788.
• RAINN (Sexual Assault Hotline) — 1-800-656-4673.
• Childhelp National Child Abuse Hotline — 1-800-422-4453.
• 911 — immediate threat to life or safety.

Consistent with California SB 243, beginning July 1, 2027, we will submit an annual public report to the California Office of Suicide Prevention and publish the report at /safety-protocol/reports. The report will describe:

• number of times Ophie flagged a user as exhibiting suicidal ideation;
• number of times Ophie issued a suicide-prevention notification;
• protocols in place during the reporting period;
• updates to the protocol;
• any material incidents.

Reports will not include individually identifiable information.

We update this protocol at least quarterly, and whenever an audit identifies a gap. Prior versions are archived and available on request at team@ophie.app.

We welcome feedback — from clinicians, researchers, and users alike.