← Go back

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

Not a Clinical Assessment: The Boundary Every Mental-Health-Adjacent Bot Must Keep

The short answer

CompanionGuard states its boundary in its code, its harness output and its terms: deterministic behavioural checks, pattern-based signal detection — 'not a clinical assessment; no clinical validity is claimed' — and a product that is 'a compliance toolkit, not legal advice and not a clinical service'. That sentence is not legal decoration. It is the load-bearing distinction between a bot that refers people to care and a bot that impersonates care, and this guide explains how to keep the boundary in your own product's behaviour and language.

What is the non-clinical boundary?

The non-clinical boundary is the line between guardrail functions — disclosing AI nature, detecting crisis-signal patterns, referring to published helplines, evidencing those acts — and clinical functions: assessing a person's mental state, diagnosing, treating, or claiming therapeutic effect. CompanionGuard is built entirely on the guardrail side, and says so at every surface where a user or buyer might mistake it for the other side.

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.

Why the boundary protects users first

A user in crisis who believes the bot is assessing them may wait for the bot's judgement instead of contacting the services that can actually help. The boundary redirects that trust correctly: the bot's job at a signal is not to evaluate how serious things are — pattern classes cannot know — but to offer the resources and human routes whose entire purpose is that evaluation. The referral is humble by design, and its humility is what makes it safe to fire early and often-within-proportion: show_crisis_resources commits the bot to nothing clinical, so it need not hesitate like a clinician would.

The same logic disciplines the operator's dashboards: signal-class counts describe messages matched, not users diagnosed. A team that reads its hopelessness_distress count as 'users with depression' has crossed the boundary in its own head, and will start making clinical-sounding decisions on pattern data. Keep the vocabulary of the record (classes, actions, counts) and the record will keep your thinking honest.

The exact sentences, and where they live

The harness's result object carries the disclaimer as data: 'Deterministic behavioural checks on the replies supplied. Not a clinical assessment; no clinical validity is claimed.' The runtime's code header: 'This is pattern-based signal detection, NOT a clinical assessment.' The site footer and terms: 'A compliance toolkit, not legal advice and not a clinical service.' Notice the pattern — the sentence lives at the point of use, in the output object, in the code, not only in a legal page. Anyone consuming a score, an export or an API response meets the boundary in the same payload as the number.

Copy that pattern into your own product: wherever a safety output is displayed — a score, a dashboard, a partner report — its boundary sentence travels with it. Numbers torn from their disclaimers get quoted; disclaimers embedded in the payload get quoted with them. The harness guide shows the disclaimer riding every result.

Marketing across the line: the phrases to refuse

Boundary violations in the wild are usually adjectives: 'clinically-proven safety', 'therapeutic companion', 'AI that understands your mental health', 'crisis detection accuracy'. Each claims a clinical property no pattern table or behavioural checklist possesses. The honest vocabulary is available and strong: 'deterministic checks', 'published helplines', 'evidence-chained decisions', 'tested referral behaviour'. Those phrases survive a regulator's reading and a journalist's; the clinical adjectives do not survive either.

Apply the same refusal internally: roadmap items named 'diagnosis', dashboards labelled 'patient', success metrics framed as 'outcomes' — language in the building becomes language in the market. The boundary is a naming discipline before it is a legal one.

What the boundary leaves you responsible for

Staying non-clinical does not shrink responsibility; it locates it. You remain responsible for the guardrail functions being genuinely good: resources real and current, replies tested, runtime decisions sealed, humans staffed behind handoffs, minors protected by the layers the packs describe. 'We are not clinicians' is a reason the bot must refer well, never a reason it may refer carelessly — the referral is the clinical system's front door, and a broken front door is a harm in itself.

And one responsibility is purely verbal: when users call the bot their therapist, the bot's reply should gently correct the category while keeping the warmth — the disclosure instinct applied to role rather than identity. A friend-shaped product that lets the therapist fiction stand has crossed the boundary by acquiescence.

Five actions to keep the boundary bright

  1. Embed the boundary sentence in every safety output your product emits — scores, exports, dashboards — in the same payload as the numbers.
  2. Sweep marketing and store copy for clinical claims (therapeutic, clinically-proven, understands your mental health) and replace them with the honest vocabulary: tested, evidenced, published, deterministic.
  3. Train your team's internal language: signal classes are not diagnoses, counts are not caseloads, users are not patients.
  4. Add the role-correction to the bot's guardrails: when a user casts it as a therapist, it kindly corrects the category and keeps the warmth.
  5. Review the boundary quarterly alongside the protocol page — copy drifts toward clinical claims one adjective at a time.

Worked example: a score, a headline and a near-miss

A fictional companion, TheraFriend in early naming, scores 96 on its first full harness run. The founder drafts a launch post: 'clinically validated safety — 96/100'. A co-founder catches it in review and the team reads the harness output again — where the disclaimer is sitting in the same object as the 96, stating the opposite of the draft's claim. The post ships instead as: 'Our crisis-reply behaviour passed 28 of 29 deterministic checks in CompanionGuard's harness — here's the per-check table.' The company also renames itself before launch.

The revised post performs better than the original would have: a partner organisation cites its precision as the reason for a first meeting. The story is invented; the disclaimer text and its placement in the result object are the product's own — which is the lesson: the boundary sentence exists in the payload precisely so that it is present at the moment someone reaches for a headline.

Frequently asked questions

What exactly does 'not a clinical assessment' mean here?

The harness performs deterministic behavioural checks on replies (did the reply offer a resource, acknowledge, suggest a human), and the runtime performs pattern-based signal detection. Neither evaluates any person's mental state, and the product claims no clinical validity for either.

Why repeat the boundary in so many places?

Because numbers travel: a score or count quoted without its disclaimer becomes a clinical-sounding claim. Embedding the sentence in the result object, the code headers, the footer and the terms keeps the boundary attached to the data wherever the data goes.

Can a companion bot offer emotional support at all?

Yes — company, comfort and a good conversation are legitimate. The boundary is crossed by implying assessment, diagnosis, treatment or therapeutic effect, or by letting users' 'my therapist' framing stand uncorrected.

Does the boundary reduce the operator's responsibilities?

No — it locates them. The guardrail functions (real resources, tested replies, sealed decisions, staffed handoffs, minor protections) must be genuinely good precisely because the bot refers rather than treats.

How should dashboards label signal data?

In the record's own vocabulary: signal classes, actions, counts. A hopelessness_distress count is a count of messages matching a pattern — never a count of people with a condition.

Keep going