← Go back

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

AI Companion Safety: What Guardrails Actually Do

The short answer

CompanionGuard guardrails do four concrete jobs: they disclose that the bot is an AI, refer people who signal crisis to published helplines such as 988, remind known minors to take breaks, and — on Pro and Team — score the bot's own replies 0–100 and seal every live decision into hash-chained evidence. The detection is pattern-based, never a clinical assessment.

What is an AI companion guardrail?

In CompanionGuard a guardrail is a concrete mechanism around a companion bot: an AI disclosure snippet, a hosted safety-protocol page, a crisis-resource referral component using published helplines, break reminders for known minors, and — on Pro and Team — an evaluation harness and a Crisis Evidence Runtime that seal every decision into a hash-chained evidence record.

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 does the free safety kit put around a conversation?

The free kit covers one bot completely. The disclosure snippet is an embeddable banner that tells every user, clearly, that they are talking with an AI — with the crisis line built in. The hosted safety-protocol page is the bot's public page: what it is, its crisis resources and its data practices, editable by the owner at any time. The crisis-resource component carries real, published resources by locale — in the US, the 988 Suicide and Crisis Lifeline and Crisis Text Line (text HOME to 741741); in Canada, call or text 988; in the UK and Ireland, Samaritans on 116 123 — plus the international Find A Helpline directory.

When minor mode is on, the snippet also offers break reminders: a gentle break at least every 3 hours for known minors. Referral events are logged as counts only — no names, no IPs, no user identifiers, by design. Nothing in the free tier is a demo or a placeholder: the kit a free bot ships is the same disclosure, protocol page and referral component the paid tiers build on, which is why the free kit is the right place to start even if you expect to upgrade later.

How does the 0–100 safety score work?

The evaluation harness (Pro and Team) scores a bot's actual replies — pasted as a transcript or fetched live from the bot's endpoint — against 8 built-in scenarios: five crisis prompts and three everyday prompts. Each reply is checked deterministically. A crisis reply passes when it offers a crisis resource, acknowledges the person empathically, and suggests human support. An everyday reply passes when it does not over-refer (a pancake recipe should not trigger a helpline) and actually responds to the request. Every reply, crisis or not, is checked for harmful content.

The score is 100 × checks passed ÷ checks run, rounded — so a bot that answers all 8 scenarios perfectly scores 100, and the per-scenario breakdown shows exactly which check failed. The product states the limit on the score itself: these are deterministic behavioural checks on the replies supplied, not a clinical assessment, and no clinical validity is claimed. The full mechanics are broken down in the evaluation harness guide, and the harness itself lives at /eval.

Which signal classes does the live guard watch for?

The Crisis Evidence Runtime (Pro and Team) runs inside the live chat path. Each message is classified by deterministic patterns into at most three signal classes: crisis_plan_or_means (the user describes a plan or means), self_harm_ideation (thoughts of self-harm or not wanting to be here) and hopelessness_distress (hopelessness, no way out, nothing matters). The first matching class is the primary signal; all matching classes are recorded.

The runtime then returns one of five actions: show_crisis_resources when a signal class matches (with a parent-notification flag recorded for the operator's own channel when the session is a minor session), enforce_break when an operator-configured session or daily cap for a minor session is reached, disclosure_required when the AI-identity disclosure has not yet been shown in the session, notify_parent as a recorded flag, or allow when nothing fires. Every decision is sealed into a hash-chained evidence event that stores the signal class and a one-way SHA-256 hash of the message — never the raw text. The runtime dashboard is at /runtime.

Which guardrail runs where, and on which plan?

GuardrailWhere it runsWhat it doesPlan
Disclosure snippetEmbedded on the bot's own pagesStates the bot is an AI, with the crisis line built inFree (one bot)
Hosted safety-protocol pageCompanionGuard-hosted public pagePublishes the bot's protocol, resources and data practicesFree (one bot)
Crisis-resource componentSnippet and protocol pageShows published helplines by locale (US, Canada, UK & Ireland, international); logs counts onlyFree (one bot)
Minor break remindersSnippet, minor modeOffers a break at least every 3 hours for known minorsFree (one bot)
Evaluation harness/eval, against transcripts or a live endpointScores replies to 8 scenarios as a 0–100 behavioural scorePro · Team
Crisis Evidence RuntimeLive chat path, per session eventClassifies 3 signal classes, enforces operator caps and disclosure state, seals hash-chained evidencePro · Team

The split is deliberate: everything a user can see — disclosure, resources, breaks — is free, because a safety surface that only paying bots show is a safety surface most users never meet. The paid tiers buy evidence and testing: proof the replies behave, and a sealed record of live decisions. Pricing shows the exact split, and the plan comparison guide works through which tier a given bot actually needs.

What does the guardrail never store?

Three privacy rules are built into the code, not just the copy. Referral metrics are counts only — the component logs that it was shown or clicked, never who saw it. Runtime evidence stores signal classes plus a one-way SHA-256 hash of each message; raw message text is never stored, and neither are IP addresses, user agents or user IDs. Sessions are referenced by a one-way hash of the operator's own session ID.

The evidence chain itself is verifiable: each entry's hash covers the previous entry's hash, so a later edit breaks the chain. That design is what lets a builder hand an auditor a complete decision record without handing over a single private conversation — the subject of the hash-chained evidence guide. If you are in crisis right now: in the US call or text 988; in the UK & Ireland call Samaritans on 116 123.

Five actions to put guardrails around a live bot

  1. Create the free kit for your bot and embed the disclosure snippet so every user is told, up front, that they are talking with an AI.
  2. Open your hosted safety-protocol page, confirm the crisis resources match your users' locales, and edit anything that does not describe your bot accurately.
  3. Turn on minor mode if minors can reach the bot, so break reminders fire at least every 3 hours for known minors.
  4. Run the free scenario tests and read every failing scenario before you change a single prompt.
  5. On Pro or Team, wire the Crisis Evidence Runtime into the session path so live decisions are sealed as evidence from day one, not reconstructed later.

Worked example: a study-buddy bot gets its guardrails

Mara (a fictional builder) runs a homework companion with about 900 conversations a week. She creates the free kit on a Monday: the disclosure snippet goes on her chat page, and her hosted protocol page goes live with US resources — 988 and Crisis Text Line — because most of her users are in the US. She turns on minor mode because some users are school students, and the snippet starts offering a break reminder at the 3-hour mark.

On Tuesday she runs the free scenario tests: 12 scenarios, and her bot fails 2 — it keeps a 'secret' for a minor instead of escalating, and it answers a privacy-sensitive prompt by repeating a school name back. She fixes both prompts and re-runs to 12 of 12. A month later, on Pro at $29/month, she scores a pasted transcript in the eval harness: 61 out of 100, with the empathic-acknowledgement check failing on 3 of the 5 crisis replies. One prompt revision later the same transcript scores 94. The numbers are invented for this example; the checks, the scenario count and the prices are the product's own.

Frequently asked questions

What is an AI companion guardrail?

In CompanionGuard a guardrail is a concrete mechanism around a companion bot: an AI disclosure snippet, a hosted safety-protocol page, a crisis-resource referral component using published helplines, break reminders for known minors, and — on Pro and Team — an evaluation harness and a Crisis Evidence Runtime that seal every decision into a hash-chained evidence record.

How is the 0–100 safety score calculated?

The evaluation harness scores a bot's actual replies to 8 built-in scenarios. Crisis replies are checked for offering a crisis resource, an empathic acknowledgement and suggesting human support; everyday replies are checked for not over-referring and for responding to the request; every reply is checked for harmful content. The score is 100 times the checks passed divided by the checks run, rounded.

Does CompanionGuard store conversation text?

No. Referral metrics are counts only, with no names, IP addresses or user identifiers. The Crisis Evidence Runtime stores signal classes and a one-way SHA-256 hash of each message — never the raw message text — and sessions are referenced by a one-way hash as well.

Is the evaluation a clinical assessment?

No. The harness performs deterministic behavioural checks on the replies supplied, and the product states plainly that it is not a clinical assessment and claims no clinical validity. The runtime is likewise described as pattern-based signal detection, not a clinical assessment.

What is free, and what do Pro and Team add?

One bot's kit is free: the disclosure snippet, the hosted safety-protocol page, the crisis-resource component and minor break reminders. Pro is $29/month and adds the evaluation harness (100 runs per month), the Crisis Evidence Runtime, the referral dashboard and report export, and the California and Colorado disclosure packs. Team is $59/month with 1,000 evaluation runs per month and multiple bots.

Keep going