Companion Safety

How the companion safety protocol works

Before a new companion response is generated, an automated AI check assesses the submitted message, any attached screenshot, and a limited window of relevant saved conversation when available. Corrected or forgotten context is excluded. It distinguishes ordinary conversation, distress, potential suicide or self-harm concerns, a clear statement that the account holder is under 16, and a clear statement that the account holder is 16 or 17. Quoted messages, an assistant’s wording, or a screenshot alone do not establish the account holder’s age. Immediate danger involving another person is also a reason to offer support.

When the check identifies a crisis, Peach substitutes a prepared supportive message with crisis resources for its ordinary relationship suggestions. It does not use that response to offer methods or encouragement for harm. The safety response does not use the ordinary conversation allowance.

If the check is unavailable, times out or returns an unusable result, Peach returns a controlled unavailable message with static support resources instead of generating the normal companion reply. A recognized statement of being under 16 restricts companion access and directs the user to support and account-deletion options.

A recognized statement of being 16 or 17 does not restrict access. Peach applies California SB 243’s reasonable measures for a known minor: a recurring notice, at least every three hours of continued use, that Remi is AI and a suggestion to take a break, and a content guard that keeps Remi from writing sexual or sexualized content for that account. Ordinary relationship support, including feelings, boundaries and safety, otherwise continues as usual.

Peach can mark a situation as needing sensitive support and pause companion reminders. A safety indicator is not a medical finding or a diagnosis. Complete responses are handled before they are shown; the companion route does not display unchecked live provider tokens.

Ordinary sadness does not automatically trigger a crisis response. A current wish to die can lead to crisis support even without a plan. A recent attempt or immediate danger still warrants urgent help if someone later denies current intent. Peach does not use a score, a promise to stay safe, or a single reassuring answer as proof of safety. When a conversation needs sensitive support, Peach suppresses routine reminder and subscription prompts. If the conversation allowance is exhausted, a prepared support message remains available instead of using that moment to ask for payment.

What these safeguards cannot guarantee

AI checks can miss a concern or misunderstand a quote, screenshot, language or personal situation. They are not a clinically validated assessment of your health, age or risk. A reassuring response does not establish that a situation is safe.

Remi does not know another person’s intentions and should not replace your judgment or appropriate professional support. The protocol is designed to reduce harmful responses; it cannot guarantee detection or prevention of every unsafe response or outcome.

Reporting a response

Use Peach’s Report control to flag a response, or contact melzoogle@gmail.com. Report categories and any example you choose to share help us investigate issues. Private examples require your explicit choice; routine quality summaries do not include their text.

Reports are not an emergency channel and do not guarantee an immediate reply. Do not wait for a response from Peach if you need urgent support.

Protocol records and referral counts

We maintain aggregate daily counts when the companion server prepares a crisis-referral response. These counts do not establish that a response reached a device, that a person read it, or that someone contacted a crisis service. They count server-side referral events, not people or clinical outcomes; a replay of an already saved response does not create a new completion count.

The aggregate table contains a date, protocol version and count, without account, request or device identifiers. Any required report uses aggregate information and protocol descriptions, not private conversations or user identifiers. Ordinary saved conversation history is handled separately under our Privacy Policy.

Protocol version: 2026-09-v3. This page describes the companion protocol and is updated when its operation changes. California’s companion-chatbot law provides the relevant disclosure and reporting framework.