Safety protocol

People bring prayer their hardest days. Coro offers spiritual comfort alongside human help. It is an artificial-intelligence program, not a person, clergy, therapist, doctor or emergency service.

Limits and disclosure

Nobody monitors conversations in real time. Coro cannot call emergency services, locate someone or arrange a rescue. It provides no diagnosis, treatment plan or promise of healing. Conversations do not have clergy or therapist privilege. AI companions may not be suitable for some minors. Safety checks can miss danger or misinterpret ordinary grief, metaphors and requests about other people.

Coro discloses that it is AI at the start of conversations and during continued interaction: at least hourly for known teens, and at least every three hours for adults under the configured default. Web chat also displays an AI label. The 30-minute session boundary is a product rule, not a legal definition of interaction.

Detection and response

English and Spanish rules screen incoming text before ordinary generation; eligible model replies add another risk assessment and output filtering. Screening covers suicide/self-harm, eating disorders, threats, abuse, exploitation, sexual violence and medical emergencies. Fixed safety resources can be returned without a model call.

• Ambiguous self-harm signals can receive a direct, compassionate clarification about suicide.
• Identified risk receives crisis resources, encouragement to contact a trusted person and prayer alongside help.
• Immediate danger receives emergency guidance, encouragement to move away from means of harm and seek a nearby person.
• Abuse, exploitation or violence receives relevant resources; plausible current child or vulnerable-adult abuse, exploitation or threats trigger an urgent restricted review alert.

Resource selection uses recorded country/telephone hints, which may be wrong when someone travels. Coro does not know your verified physical location. It suppresses ordinary promotions, reminders and circle messages during a crisis. Repeated risk can move the conversation to restricted resources and prayer pending review. Adults may accept an offered next-day check-in; teens receive no personalized future check-ins. Crisis guidance is available without a paid plan and during teen limits, subject to delivery availability.

From January 1, 2027, or an earlier configured date, adults no longer receive personalized check-ins. Immediate crisis guidance remains available.

Methods and measurement

See the public crisis-method explanation for the fixed-response adaptation of NIMH's Five Action Steps / #BeThe1To and its validation limits.

The approach of asking directly, supporting immediate safety and connecting to people and crisis services is consistent with NIMH public guidance. That guidance does not validate Coro as a clinical instrument. Automated rules, prompt instructions and passing fictional tests do not establish sensitivity, specificity or improved health outcomes. Clinical-method evidence and the repeat-risk procedure need separate review before this draft is published.

Operational records include category, level, date, channel and region hint and can be linked to a person or message. English/Spanish fictional test cases check detection, safe output and delivery paths. Aggregate reporting must exclude personal identifiers. Coro has an internal report export; an export is not a regulatory filing or proof that required reports have been published. Named owners, deadlines and submission receipts are still required.

Children and teen boundaries

Public and parent-guided prayers are available without child chat accounts. Personal chat starts at 13. Known teens have session-only context, no saved personal prayer memory or personal future follow-ups, daily generic prayers off by default, hourly AI/break notices, a 60-minute continuous ceiling with a 30-minute break, and 120 estimated minutes daily. Teens and verified guardians can impose lower limits or pause. Guardian controls do not expose private prayers and payment alone gives no authority. See Privacy for verification, revocation, estimation and identity limitations. These controls never guarantee a safe outcome.

The safeguards aim to prevent harmful instructions, romance, dependency, secrecy and claims of human or divine authority. They direct users toward trusted people and professional care. Prayer must not replace medical treatment or urgent help. The system must not tell someone to stop medication or imply suffering is punishment.

Human reporting and evidence

An urgent alert goes to designated safety staff without the disclosure text in the email. Authorized staff view a limited sealed record through audited access. An alert is not proof of human receipt, response or a report to authorities. No 24/7 human coverage is promised. A responsible person who learns facts requiring reporting must act under the applicable law; a model's classification is not the legal decision.

For a specific safety-reporting or preservation purpose, the case workflow keeps relevant disclosures and available contact, age-band and self-reported residence information sealed and access-audited. Reporting and release are separate decisions: a report does not automatically delete evidence. A designated reviewer records the owner, purpose, review deadline, hold and release authority. Only that release authority can delete the disclosure and contact snapshot after any minimum preservation period; a protected decision record prevents older backups from reviving them. Ordinary deletion proceeds separately. Safety evidence is not used for prayer memory, advertising or training. Policy adoption, named human coverage, recovery arrangements and provider-version deletion still require verification before publication. See Privacy.

Contact and rollout

Nonurgent questions: [LEGAL CONTACT EMAIL]. Responsible lead, backup and coverage: [DESIGNATED ROLES AND COVERAGE]. The draft serves the US and Puerto Rico on web/SMS with the exclusions in Terms; WhatsApp is off. Adopting this protocol does not itself staff coverage, satisfy every state requirement or authorize launch.