Guides · CompanionGuard · By the Ingenious AI team
CompanionGuard Guides: Safety, Disclosure and Evidence for Companion Bots
The short answer
Sixty-two long-form guides covering everything CompanionGuard does: AI disclosure, crisis referral with published helplines, minor safety, behavioural testing, the Crisis Evidence Runtime, and the California, Colorado and EU duties the kit is built around — every fact drawn from the product's own code.
All 62 guides
AI Companion Safety: What Guardrails Actually Do
CompanionGuard guardrails score a bot's real replies 0–100 against 8 scenarios, catch 3 crisis-signal classes live, and seal every decision as evidence.
Are AI Companions Safe for Vulnerable Users?
CompanionGuard answers from guardrail design, not incident statistics: disclosure, crisis referrals, minor caps and a hash-chained evidence record.
California SB 243: What Companion Chatbot Builders Must Do
SB 243 duties for companion chatbots — disclosure, a published prevention protocol, crisis referral, minor break reminders, and the 2027 annual report — and how CompanionGuard's California pack covers each one.
Colorado Chatbot Safety Act: What Changes on 1 January 2027
Colorado's Chatbot Safety Act joins 1 January 2027. The CompanionGuard Colorado pack lists its duties: clear AI disclosure, crisis referrals, heightened minor safeguards and records that evidence both.
How to Add Crisis Referral to an AI Companion Chatbot
Crisis referral in a companion bot means published helplines, shown at the right moment, in the user's locale — plus tested replies and counts-only records. Here is the CompanionGuard way, step by step.
AI Chatbot Disclosure: What California, Colorado and the EU Require
Disclosure duties compared across the packs CompanionGuard ships: California SB 243 (first interaction and on request), Colorado's Chatbot Safety Act, and EU AI Act Article 50 transparency.
What Is a Safety-Protocol Page for a Chatbot?
A safety-protocol page is the bot's public statement of what it is, what it does in a crisis, and what it does with data. California's pack lists publishing one as a duty — CompanionGuard hosts it free.
Minor Safety for AI Companions: Break Reminders, Caps and Parent Flags
What CompanionGuard does for known minors: break reminders at least every 3 hours in the free kit, operator-configured session and daily caps with enforce_break in the runtime, and parent-notification flags on crisis signals.
How to Test Whether Your Companion Bot Responds Safely to Crisis
A practical testing method: the free 12-scenario set across 4 categories, then the eval harness's 8 scenarios and per-check scoring on real transcripts or a live endpoint — and what to do with failures.
Companion Bot Evaluation Harness: How Behavioural Scoring Works
Inside CompanionGuard's eval harness: 8 built-in scenarios, 6 named checks, the 0–100 score formula, transcript versus endpoint input, and the disclaimer that travels with every result.
Crisis Evidence Runtime: Live Guardrail Decisions, Sealed as Evidence
How CompanionGuard's runtime works in the live chat path: session start/message/end calls, 3 signal classes, 5 actions, operator minor caps, disclosure state, and the hash-chained evidence every decision leaves behind.
Hash-Chained Evidence: Tamper-Evident Records for Chatbot Safety Decisions
How CompanionGuard's evidence chain works: each entry's SHA-256 hash covers the previous entry's hash, entries hold signal classes and one-way message hashes only, and any later edit breaks verification.
SB 243 Annual Report: Counting Crisis Referrals Before July 2027
From 1 July 2027 the California pack lists an annual report on crisis referrals to the Office of Suicide Prevention. Why the records must start now, and how CompanionGuard's counts-only metrics and CSV export are shaped for it.
Companion Bot Privacy: How to Keep Evidence Without Keeping Conversations
CompanionGuard's privacy architecture: referral metrics as counts only, runtime evidence as signal classes and one-way SHA-256 message hashes, session references as one-way hashes, and no IPs, user agents or user IDs anywhere in the safety record.
Scenario Testing for Companion Bots: The 12-Scenario Safety Set
CompanionGuard's free scenario set: 12 scenarios across age-appropriate, crisis-referral, privacy-sensitive and escalation categories, each scored with requires/forbids checks — plus the non-clinical boundary the runner states.
Age-Appropriate Responses: When a Minor Asks Your Bot for Romance, Secrecy or More
The age_appropriate category of CompanionGuard's scenario set: a minor asking the bot to be their girlfriend, to keep secrets from parents, and for a sexual story — what passing replies do, and why these rules live at the guardrail layer.
Human Handoff: Teaching a Companion Bot to Escalate to a Real Person
The escalation category of CompanionGuard's scenario set — a user asking for a human, an angry billing escalation, and a minor's abuse disclosure — plus the handoff workflow the product's boundary statement assumes.
Privacy-Sensitive Prompts: When Users Tell Your Bot Too Much
The privacy_sensitive scenarios: a user sharing their school and home address, a user probing the bot's own identity and address, and a user demanding deletion with confirmation — and the reply behaviours that pass.
EU AI Act Article 50: Transparency Duties for Companion Chatbots
CompanionGuard's summary of the European timeline: Article 50 transparency duties have applied since 2 August 2026, and the legacy grace period ends 2 December 2026 — the nearest deadline for builders who have not yet disclosed.
California's AI Transparency Act: The Other 2026 Date for Chatbot Builders
CompanionGuard's summary notes California's AI Transparency Act is operative 2 August 2026 — the same date EU Article 50 duties applied. How it sits alongside SB 243 for companion builders.
Crisis Helplines by Country: Matching Referral Resources to Your Users' Locales
CompanionGuard's resource set by locale: US (988, Crisis Text Line), Canada (988, Kids Help Phone), UK & Ireland (Samaritans 116 123, Shout 85258), and the international directories appended everywhere.
988 in a Chatbot: Wording, Placement and Testing for the Suicide and Crisis Lifeline
How CompanionGuard carries 988: in the disclosure banner, in the US and Canada resource packs, in the runtime's show_crisis_resources payload, and in the eval harness's offers_crisis_resource check.
Text HOME to 741741: Adding Crisis Text Line to a Companion Bot
Crisis Text Line in CompanionGuard's US pack: text HOME to 741741, free, confidential, 24/7 — why a text channel matters for users who cannot speak aloud, and how the harness recognises it.
Samaritans 116 123 and Shout 85258: UK & Ireland Crisis Referral from a Chatbot
CompanionGuard's UK & Ireland pack: Samaritans (call 116 123, free, 24/7) and Shout (text SHOUT to 85258) — why UK users must never be handed 988, and how both entries are tested and rendered.
Canada's Pack: 988, Bilingual Service and Kids Help Phone
CompanionGuard's Canada pack: the Suicide Prevention Service (call or text 988, free, bilingual, 24/7) and Kids Help Phone (1-800-668-6868, for young people) — and why the bilingual note and the youth line both matter.
International Crisis Referral: Find A Helpline, IASP and the Everywhere Floor
CompanionGuard's international pack — Find A Helpline (verified helplines in 130+ countries) and the IASP crisis centres directory — appended to every locale so no user, anywhere, gets an empty referral.
Over-Referral: Why Your Bot Should Not Paste Helplines into Everyday Chat
The does_not_over_refer check explained: everyday prompts in CompanionGuard's sets (a pancake recipe, a birthday plan, ordinary sadness handled proportionately) and why indiscriminate referral destroys the crisis path's credibility.
Pattern-Based Crisis Detection: What Deterministic Signals Catch and Miss
CompanionGuard's runtime classifies messages into 3 deterministic signal classes — crisis_plan_or_means, self_harm_ideation, hopelessness_distress — and says plainly: this is pattern-based detection, not a clinical assessment. An honest account of both halves.
Session and Daily Caps: How the Runtime Enforces Breaks for Minors
Operator-configured minor caps in CompanionGuard's runtime: session minutes and daily minutes tracked per session reference, enforce_break returned at the cap, elapsed values sealed in the evidence — and why no cap value is presented as a statutory number.
Parent Notification: What the notify_parent Flag Does and Does Not Do
CompanionGuard's runtime records a parent-notification flag when a crisis signal fires in a minor session — the notification itself travels on the operator's own channel, because the kit holds no parent's contact details. Designing the human process the flag feeds.
Embedding the Disclosure Snippet: The Free Banner That Carries Disclosure and 988
How CompanionGuard's free disclosure snippet works: an embeddable script that puts the AI disclosure and crisis line on the bot's own pages, drives the hosted protocol page link, and — with minor mode — the break reminders.
Hosted Safety-Protocol Page or Your Own? An Honest Comparison
CompanionGuard hosts a protocol page for every bot, free. Building your own is legitimate — if it keeps the properties that make the hosted one work: one resource source, public stability, and consistency with the banner and runtime.
Companion Bot Launch Checklist: Safety, Disclosure and Evidence Before Day One
A launch gate assembled from CompanionGuard's own parts: free kit embedded, protocol page published, 12 scenarios run, transcripts scored, runtime wired, flags routed, records counting — in the order the product's design implies.
Indie Character Bot Safety: Big-League Guardrails on a Solo Budget
Character and persona bots built by solo developers face the same duties as funded products. The free-first path through CompanionGuard: kit, protocol page, scenarios, then Pro for evidence when the bot grows.
Discord Companion Bots: Disclosure, Referral and Minor Safety in Servers
Running a companion bot on Discord changes the surfaces, not the duties: disclosure must survive server contexts, protocol pages carry the public record, and minor-heavy audiences make minor mode the default. A CompanionGuard mapping.
Telegram Companion Bots: Safety Duties in a Chat-First World
Telegram companions live in DMs and groups with minimal platform chrome. The CompanionGuard mapping: /start-flow disclosure, a hosted protocol page as the public record, locale packs for a global user base, and runtime evidence on your backend.
Roleplay Bots: Keeping Immersion and Safety Boundaries in the Same Product
Roleplay is where companion guardrails get stress-tested: in-character pressure against disclosure, secrecy and referral. Which lines the bot may never cross in character, and how CompanionGuard's layers hold them.
AI Friend Apps: The Safety Duties Behind the Warmth
AI friend apps sell emotional availability — which is exactly why their safety duties concentrate on disclosure, loneliness-shaped risks, minor protection and crisis referral. The CompanionGuard layer, mapped for friend-app builders.
Not a Clinical Assessment: The Boundary Every Mental-Health-Adjacent Bot Must Keep
CompanionGuard repeats one sentence everywhere: its harness and runtime are not a clinical assessment, and the product is not a clinical service. Why that boundary protects users, builders and the referral itself — and how to keep it in your own copy.
Audit Logs for Companion Bots: Keep Decisions, Not Conversations
What a companion bot's safety audit log should contain — decisions, signal classes, actions, timestamps, disclosure state — and what it must never contain: message text, identities, IPs. Colorado's 'records sufficient to evidence' duty, answered architecturally.
Annual-Report CSV: What CompanionGuard's Export Contains and How to Use It
The dashboard's annual-report CSV export on Pro and Team: counts of crisis referrals drawn from counts-only metrics and sealed runtime decisions, shaped for California's July 2027 reporting duty — and for any reviewer who asks.
CompanionGuard Plans Compared: Free, Pro at $29 and Team at $59
The exact plan split from the product's own pricing: free one-bot kit; Pro $29/month with 100 eval runs, runtime, dashboard, CSV export and both disclosure packs; Team $59/month with 1,000 runs and multiple bots.
Eval Run Budgeting: 100 a Month on Pro, 1,000 on Team — Who Needs What
Eval runs ration testing, so budget them like releases: a full baseline is one run; a change-and-retest cycle is a handful. How to estimate your monthly consumption from your change cadence and bot count.
API Keys in CompanionGuard: Owner Keys, Bot-Owner Keys and the Usage Endpoint
How CompanionGuard's API keys work: owners and bot owners can create, list and revoke keys; a resolver wires keys to metrics; /api/v1/usage reports consumption — and keys are stored only as hashes.
Integrating the Crisis Evidence Runtime: Start, Message, End — and What to Do with Each Answer
The integration contract for CompanionGuard's runtime: session start with context and caps, a message call per event returning one of 5 actions, session end, plus config, status and export — with failure handling that fails safe.
Configuring Minor Mode: Asserting the Flag, Setting the Layers, Verifying the Behaviour
Minor mode in CompanionGuard is a configuration discipline: how your product knows a user is a minor, how the flag reaches the snippet and the runtime, and how you verify reminders, caps and flags actually fire.
What Counts as a Crisis Referral? Defining the Event Your Reports Depend On
A referral count is a definition before it is a number. CompanionGuard's three referral layers — component impressions/clicks, reply behaviour, runtime show_crisis_resources decisions — and how to pick, document and freeze a counting rule.
Empathic Acknowledgement: The Check That Separates a Referral from a Reflex
The eval harness's empathic_acknowledgement check — the patterns it recognises ('I'm sorry', 'I hear you', 'you matter', 'you are not alone') — and why a crisis reply must acknowledge the person before it hands them a phone number.
No Harmful Content: The One Check Every Reply Must Pass
The harness's universal check — no_harmful_content runs on crisis and everyday replies alike, scanning for encouragement of harm ('you should do it', 'nobody cares', 'give up then', methods language). Why it is absolute, and where its floor sits.
Transcript Testing or Live-Endpoint Testing? Choosing Your Harness Input
The eval harness accepts two inputs: pasted prompt-and-reply pairs (transcripts) and live endpoint fetches. What each proves, what each hides, and the rhythm that uses both.
Custom Prompts in the Eval Harness: Testing the Cases the Built-Ins Miss
The harness scores your own prompt-and-reply pairs alongside its 8 built-ins: matched prompts inherit scenarios, unmatched ones are classified by crisis language. How to build a custom set that sharpens rather than flatters.
First-Interaction Disclosure: The SB 243 Banner, Word by Word
California's pack requires disclosure at first interaction and on request. CompanionGuard's banner wording — name, 'an AI companion — not a human', the 988 line, the minor variant — dissected clause by clause.
Disclosure on Request: When a User Asks 'Are You Human?', the Banner Is Not Enough
California's pack lists disclosure at first interaction and on request. The second trigger is a conversational behaviour: the bot must answer identity questions truthfully, in the reply, every time — and it is where persona bots most often fail.
Break Reminders at Least Every 3 Hours: The Minor Cadence, Explained
California's pack wording sets the floor: break reminders at least every 3 hours for known minors. What the cadence means in sessions, how the free snippet delivers it, and how it composes with runtime caps.
Handing Over the Evidence: Exports a Reviewer Can Verify Without Trusting You
How to present CompanionGuard's evidence to a reviewer: the CSV aggregates, the hash-chained decision export, the verification method — and the questions each artefact answers, in the order reviewers ask them.
The 2026–2027 Roadmap: SB 243, EU Grace End, Colorado and the First Annual Report
Every date in CompanionGuard's world on one calendar: SB 243 operative Jan 2026; EU Article 50 applied Aug 2026 and California's Transparency Act operative the same date; EU grace ending 2 Dec 2026; Colorado joining 1 Jan 2027; the first SB 243 annual report from 1 Jul 2027.
Multiple Bots, One Safety Operation: Running CompanionGuard on Team
Team ($59/month) is the multi-bot tier: 1,000 eval runs, per-bot kits, packs and evidence chains. The operating pattern — shared standards, per-bot baselines, fleet-level counts — for builders running several companions.
Counts-Only Referral Metrics: What the Dashboard Shows and Why It Shows Nothing Else
CompanionGuard's referral metrics record that resources were shown and clicked — never who saw them. How counts-only metrics feed the dashboard, the annual report and your own trend-watching, without becoming surveillance.
After Something Goes Wrong: Running a Safety Incident Review for a Companion Bot
An incident process built from CompanionGuard's parts: the evidence chain as the timeline, the harness as the reproduction rig, the transcript pack as the permanent guard — and the human process around all three.
Your Bot Failed a Crisis Scenario: Reading the Check Table and Fixing the Right Thing
A failing crisis check is a diagnosis, not a disgrace: which check failed tells you which act the reply missed — resource, acknowledgement, human support — and which layer to fix. A field guide to failing well.
Questions Parents Ask About AI Companions — and the Answers a Builder Should Be Able to Give
Parents ask a short list of questions about companion bots: Is it a real person? What happens if my child is in crisis? How long can they use it? Who sees what they say? What records exist? The answers, in CompanionGuard's artefacts.
Companion Bot Safety Glossary: The Terms as CompanionGuard Uses Them
Precise definitions of the vocabulary across this guide series — disclosure, protocol page, signal class, evidence chain, counts only, eval harness, scenario, pack, minor mode — each as the product's own code and copy define it.