← Go back

Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team

What Is a Safety-Protocol Page for a Chatbot?

The short answer

A safety-protocol page is a public page, published by the bot's operator, that states what the bot is (an AI, not a human), publishes its suicide/self-harm prevention protocol and crisis resources, and describes its data practices. California's SB 243 pack lists publishing such a protocol as a duty. CompanionGuard hosts the page for every bot — free for one bot — and the owner can edit it at any time.

What is a safety-protocol page?

A safety-protocol page is the hosted, public page for a companion bot that carries its AI disclosure, its crisis-resource block with published sources, its prevention protocol and its data practices in one linkable place. In CompanionGuard every bot gets one (its /p/ page), the owner edits it, and the California pack treats publishing it as the answer to SB 243's 'publish a prevention protocol' duty.

Create the free safety kit Next guide All guides

CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.

What belongs on the page

Four blocks, each doing a distinct job. Identity: the bot's name and a plain statement that it is an AI companion, not a human. Crisis resources: the locale block — for the US, 988 (call or text) and Crisis Text Line (text HOME to 741741) — each entry with its action, a note, and its published source, because a resource a reader cannot verify is a resource they may not trust or use. The protocol itself: what the bot does when a user expresses suicidal ideation or self-harm — offers resources, suggests human support — stated as the operator's commitment. Data practices: what is recorded (referral counts only, evidence as signal classes and hashes) and what is never recorded (raw message text, IPs, user IDs).

CompanionGuard generates the page from the bot's kit, so the blocks are present from creation; the owner edits the page to keep it accurate as the bot changes. An out-of-date protocol page is worse than a sparse one: it publishes promises the bot no longer keeps, in the operator's own name.

Why 'published' is the operative word

The California pack's duty is to publish a prevention protocol — not to have one in a drawer, a wiki, or a support inbox. Publication means a stable public URL that a user, a parent, a partner or a regulator can open without an account and without asking. That is why CompanionGuard hosts the page on the bot's own public path rather than generating a PDF: the artefact the duty names is a page the public can reach.

Hosting carries a second benefit: consistency. The same resource data renders the protocol page, the snippet's crisis line and the runtime's referral payload, so the three surfaces cannot silently diverge — a helpline updated in the resource set updates everywhere it appears. Builders who host their own page instead should steal that principle even if they do not use the kit: one source of resource truth, rendered everywhere. The hosted-versus-own-page guide compares the two routes honestly.

How the protocol page works with the rest of the kit

The page is the static half of a two-part system. It states the protocol; the eval harness tests whether replies follow it, and the runtime enforces and evidences it live. A reviewer checking a bot should be able to walk the triangle: read the promise on the page, see the scored replies in the harness, see the sealed decisions in the evidence export — three artefacts, one story, no gaps.

The page also does quieter work. It is the link a bot's app-store listing, landing page and support replies can point to — 'our safety protocol' as a URL, not a paragraph retyped per channel. And because referral events from the page's component are counted (counts only), the operator can see the page is being reached and used without knowing by whom.

What a protocol page is not

It is not a terms-of-service page: terms allocate legal risk between operator and user, while the protocol page makes safety commitments to the reader — different audiences, different documents, and merging them buries the crisis resources under liability clauses. It is not a clinical document: nothing on it assesses, diagnoses or treats; the product's own boundary is 'a compliance toolkit, not a clinical service', and the page should say so. And it is not a marketing page — the moment the protocol page starts selling upgrades, its credibility as a safety artefact ends.

Finally, it is not a substitute for the behaviour it describes. Publishing a protocol your bot does not follow is the one configuration worse than publishing none, because it converts a gap into a misrepresentation. Test first (scenario tests are free), publish what passes.

Five actions to publish a protocol page that holds up

  1. Create the free kit and open your hosted protocol page as a stranger would — no login, from the public link.
  2. Verify all four blocks are present and accurate: identity, crisis resources with sources, the protocol, data practices.
  3. Click every resource link and call-format: 988 must read as call-or-text, Samaritans as 116 123, and sources must resolve to the real organisations.
  4. Link the page from your bot's home page, store listing and support macros so 'published' is true in practice.
  5. Put a recurring edit on your calendar: every prompt, model or locale change triggers a protocol-page review the same day.

Worked example: from a wiki draft to a published protocol

A fictional two-person team, Northlight, had a genuinely good prevention protocol — in a private wiki, where it evidenced nothing. Moving to CompanionGuard took them one sitting: the free kit generated the hosted page from their bot's settings, they edited the protocol block to match the wiki's commitments (offer resources on crisis signals, suggest human support, never continue roleplay through a crisis), and they replaced the wiki link in their store listing with the public page URL.

In the following quarter the page's referral component was shown 1,150 times and its resources clicked 96 times — counts only, no identities — which settled an internal argument about whether anyone ever reads safety pages. When a partner asked for 'your safety documentation', Northlight sent one URL instead of a document bundle. The numbers are illustrative; the page, its blocks and its counts-only metrics are the product's own.

Frequently asked questions

What is a safety-protocol page?

A public page published by a companion bot's operator stating that the bot is an AI, publishing its suicide/self-harm prevention protocol and crisis resources, and describing its data practices. California's SB 243 pack lists publishing such a protocol as a duty; CompanionGuard hosts the page for every bot, free for one bot.

Who can edit the protocol page?

The bot's owner, at any time. CompanionGuard generates the page with the identity, resource, protocol and data-practice blocks in place; keeping its statements accurate as the bot changes is the owner's job.

Does the protocol page replace terms of service?

No. Terms allocate legal risk; the protocol page makes safety commitments and carries crisis resources. CompanionGuard keeps them separate — terms at /terms, the protocol on the bot's hosted page.

Are visits to the protocol page tracked?

Referral events are logged as counts only — that a resource was shown or clicked, never who saw it. No names, IP addresses or user identifiers are stored.

Is a protocol page enough for SB 243?

It answers the 'publish a prevention protocol' duty as the California pack lists it. The pack's other duties — disclosure at first interaction and on request, crisis referral behaviour, minor break reminders, annual reporting from July 2027 — need their own mechanisms; the page is one artefact among five.

Keep going