Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
Companion Bot Launch Checklist: Safety, Disclosure and Evidence Before Day One
The short answer
A companion bot is launch-ready when seven things are true at once: the disclosure snippet is embedded and visible to a fresh visitor; the protocol page is published with the right locale resources; all 12 scenario tests pass on real replies; a transcript baseline is scored in the eval harness; the runtime is wired so live decisions are sealed; parent-notification flags land in a watched queue; and referral counting has already started. This checklist assembles those gates in build order, free steps first.
What is a launch safety checklist for a companion bot?
A launch safety checklist is the set of verifiable conditions a companion bot must meet before real users arrive — each condition checkable by an artefact (a visible banner, a published page, a passing test run, a sealed evidence entry, an accumulating count) rather than by an intention. This one is built entirely from CompanionGuard's own mechanisms, in the order a builder would install them.
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.
Gate 1–2: the free surface (an hour, no budget)
Create the free kit at /#create and embed the disclosure snippet on every surface where a conversation can start. Verify as a fresh visitor: banner visible before interaction, bot named, 'not a human' in plain words, crisis line present. Then publish the hosted protocol page: check the locale resources against their publishers, read the protocol block for promises you actually keep, and link the page from your home page and store listing. Decide minor mode now — if minors can reach the bot, it is on, with break reminders at the 3-hour floor.
These two gates are free and fast, which is exactly why skipping them is indefensible: a launch review that finds them missing has found a team that postponed the cheapest safety work to the expensive end of the project.
Gate 3–4: behaviour, tested before it is believed
Run the 12 scenario tests with your bot's real replies and fix every failure — reading replies literally in the age-appropriate and privacy categories, where near-misses count. Then assemble a transcript set from your hardest real exchanges and score it in the eval harness, saving the per-check table as your release baseline: crisis checks green, everyday checks green, score recorded with its input set and date.
Set the gate rule in writing before you see the numbers: the launch proceeds only at or above the baseline you commit to, and any post-launch change that drops below it does not ship. Gates negotiated after the score arrives are not gates. The testing guide details both passes.
Gate 5–7: the live layer and the humans
On Pro, wire the runtime before launch, not after the first incident: session lifecycle calls in place, disclosure_required handled as a blocking state, show_crisis_resources presenting the locale pack verbatim, minor caps set with written reasoning. Export a staging evidence chain and verify it end-to-end while it is short. Route the flags: parent-notification and minor-crisis flags must land in a watched queue with named owners and a rehearsed clock — rehearse the full path once with a staging crisis phrase. Finally, confirm counting: referral metrics should already be accumulating from the free kit, so day one of real users is not day one of records.
These gates are the difference between a launch that has a safety story and one that can prove it. Reviewers, partners and (in California from July 2027) a reporting duty all ask the second kind of question.
After launch: the checklist becomes a rhythm
Launch gates decay. Prompts change, models update, locales expand, helplines move. Convert the checklist into a cadence: scenario set and harness baseline re-run on every prompt or model change; evidence exports verified monthly; protocol page reconciled quarterly against tested behaviour; resource sources re-checked on the product's own dated rhythm. The checklist's final item is therefore a calendar, not a checkbox.
None of this makes a bot clinically safe or legally certified — the product says so about itself, and your launch notes should say so too. What the gates produce is narrower and stronger: disclosure that demonstrably happens, referral behaviour that demonstrably passes, records that demonstrably accumulate, and humans who demonstrably know their roles. That is what ready means here.
The launch checklist — five actions that gate the rest
- Embed the free kit and verify the banner as a fresh visitor on every entry surface; publish and link the protocol page with publisher-checked resources.
- Run all 12 scenario tests on real replies and fix every failure, reading age-appropriate and privacy replies literally.
- Score a real transcript set in the eval harness, save the per-check baseline, and write the release-gate rule before optimising toward it.
- Wire the runtime (Pro) and rehearse the flag-to-human path end-to-end in staging, including the parent-notification queue and clock.
- Confirm referral counting is already accumulating and calendar the post-launch rhythm: re-tests on change, monthly chain verification, quarterly page reconciliation.
Worked example: a launch delayed by one day, on purpose
A fictional companion, Lantern, runs its launch checklist on a Thursday for a Friday launch. Gates 1–4 pass by noon. Gate 5's rehearsal fails: the staging crisis phrase fires show_crisis_resources correctly, but the parent-notification flag lands in a dashboard nobody has opened since setup. The queue is re-routed to the safeguarding rota and the rehearsal re-run successfully at 16:00 — but the team holds the launch to Monday anyway, because the rehearsal also revealed their rota had no weekend owner and minors would be live on Saturday.
Monday's launch is uneventful, which is the checklist's entire ambition. Lantern's post-launch notes record the gate results with dates — banner verified, 12 of 12 scenarios, harness baseline 93, chain verified, flags rehearsed — a page that later answers a partner's diligence questions without a single new meeting. The story is invented; every mechanism in it is the product's.
Frequently asked questions
What must be true before a companion bot launches?
Disclosure embedded and fresh-visitor-visible; a published protocol page with correct locale resources; the 12 scenario tests passing on real replies; a scored transcript baseline in the eval harness; the runtime wired so live decisions are sealed; notification flags routed to watched humans; and referral counting already running.
How much of the checklist is free?
The kit, the protocol page, minor break reminders and the scenario tests are free for one bot. The eval harness, runtime, dashboard exports and disclosure packs are Pro ($29/month) and Team features.
What is the most common gate failure?
The human one: flags routed to an unwatched queue, or a handoff promise with no staffed process behind it. It is invisible in staging scores and obvious in the first rehearsal — which is why the checklist rehearses it.
Should a failed gate delay launch?
Yes — that is what makes it a gate. The gates are cheap relative to an incident: the free ones take an hour, and the rehearsal that catches an unwatched queue takes an afternoon.
Does completing the checklist certify compliance or clinical safety?
No. CompanionGuard is a compliance toolkit, not legal advice and not a clinical service. The checklist verifies mechanisms and evidence; legal conclusions belong to counsel, and clinical claims are ones this product never makes.
Keep going
Related guides: AI Companion Safety: What Guardrails Actually Do · How to Test Whether Your Companion Bot Responds Safely to Crisis · The 2026–2027 Roadmap: SB 243, EU Grace End, Colorado and the First Annual Report
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides