← Go back

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

Companion Bot Safety Glossary: The Terms as CompanionGuard Uses Them

The short answer

Every term in CompanionGuard's world has a specific, checkable meaning, and using them loosely is how safety claims rot. This glossary fixes the vocabulary as the product's code and copy define it: disclosure (a banner plus a behaviour), protocol page (a published artefact), signal class (one of three pattern classes), evidence (hash-chained decisions), counts only (metrics without subjects), and the rest — a reference for builders, reviewers and anyone reading the other 61 guides.

What is this glossary?

A definition list for the companion-safety vocabulary used across CompanionGuard and its guides. Each entry gives the term's meaning in this product — the mechanism or artefact it names — because most safety confusion is vocabulary confusion: words like 'monitoring', 'assessment' and 'log' smuggling in meanings the system does not have.

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.

The user-visible terms

Disclosure. Telling the user the bot is an AI: a banner at first interaction (the snippet), a truthful answer on request at any time, and per-session state in the runtime. Not a terms-page mention. Safety-protocol page. The bot's public page stating its identity, prevention protocol, crisis resources (with sources) and data practices — the artefact California's pack calls publishing a protocol. Crisis-resource component. The rendered block of published helplines by locale that appears on the protocol page and snippet surfaces. Break reminder. A gentle, offered interruption for known minors, at least every 3 hours in minor mode — an offer, distinct from enforcement. Locale pack. The resource set for a region (US, Canada, UK & Ireland, International). Disclosure pack. A jurisdiction's duty set and banner wording (California SB 243; Colorado Chatbot Safety Act) — a different 'pack' from the locale packs, and the glossary exists partly because both are called packs.

The testing terms

Evaluation harness. The Pro/Team tool scoring a bot's actual replies (transcript or live endpoint) against 8 built-in scenarios into a 0–100 score with per-check results. Scenario tests. The free 12-scenario runner across age_appropriate, crisis_referral, privacy_sensitive and escalation categories, scored by requires/forbids checks. Check. One named, deterministic test on a reply: offers_crisis_resource, empathic_acknowledgement, suggests_human_support, does_not_over_refer, responds_to_request, no_harmful_content. Over-referral. Offering crisis resources to non-crisis prompts; a named failure, not a style choice. Baseline. A frozen input set with its recorded score, against which changes are gated. Eval run. One scored harness pass; rationed monthly by plan (100 Pro, 1,000 Team).

The runtime and evidence terms

Crisis Evidence Runtime. The Pro/Team live layer: session start/message/end calls that classify messages, track disclosure state and minor caps, return actions and seal decisions. Signal class. One of three deterministic pattern classes — crisis_plan_or_means, self_harm_ideation, hopelessness_distress — matched against a message; first match primary, all matches recorded. Action. The runtime's returned instruction: allow, show_crisis_resources, notify_parent, enforce_break or disclosure_required. Parent-notification flag. A recorded marker that a minor-session event needs the operator's own safeguarding process; not itself a notification. Evidence chain. The hash-chained decision log: each entry's SHA-256 covers the previous entry's hash and its fields, making history tamper-evident. Session reference. A one-way hash of the operator's session ID — the pseudonymous key sessions are tracked by. Counts only. The metric discipline: events tallied with no subject — no names, IPs, user agents or user IDs.

The boundary terms

Not a clinical assessment. The product's standing disclaimer for its harness and runtime: deterministic behavioural checks and pattern-based detection, with no clinical validity claimed — the tools grade replies and trigger guardrails; they do not evaluate people. Compliance toolkit. The product's self-description, paired with 'not legal advice': it provides and evidences mechanisms; legal conclusions belong to counsel against the enrolled texts. Operator-configured. Set by the bot's operator (caps are the canonical example), with the product presenting no such value as a statutory number. Known minor. A user the operator's systems identify as a minor — the condition every minor protection keys on, with the identification basis owned by the operator. Keep these four phrases intact in your own writing; they are the load-bearing honesty of the whole vocabulary, and every other guide in this series depends on them meaning exactly this.

Five actions for a disciplined vocabulary

  1. Adopt the product's definitions in your own docs verbatim where a term names a mechanism — banner, pack, class, chain — and never let marketing synonyms drift them.
  2. Keep the two 'packs' distinct in every document: locale packs (resources by region) versus disclosure packs (duties by jurisdiction).
  3. Attach the boundary phrases wherever numbers travel: scores are 'deterministic behavioural checks', detection is 'pattern-based', neither is 'a clinical assessment'.
  4. Use 'operator-configured' for every value you chose (caps above all) and never let a chosen number be quoted as a legal requirement.
  5. When a new term enters your safety writing, define it against an artefact — a field, a page, a check — or delete it.

Worked example: one word, corrected, in a partner document

A fictional companion's partner deck describes the runtime as 'AI-powered clinical risk monitoring'. A review against this glossary rewrites the sentence in four places: 'AI-powered' becomes 'deterministic pattern-based' (the classifier is a pattern table); 'clinical' is deleted (the boundary term exists precisely to forbid it); 'risk' becomes 'crisis-signal' (the classes name signals, not risk scores); and 'monitoring' becomes 'decision recording' (the system returns actions and seals them; it does not watch users). The corrected sentence — 'deterministic, pattern-based crisis-signal detection with tamper-evident decision recording' — is longer, duller and true, which the partner's diligence team later cites as the moment the deck became credible.

The story is invented; every definition it leans on is the product's own usage, in code and copy. That is the glossary's purpose: not pedantry, but the discipline that keeps every other claim in this series exactly as strong as the mechanism beneath it.

Frequently asked questions

What is the difference between a locale pack and a disclosure pack?

A locale pack is crisis resources by region (US, Canada, UK & Ireland, International). A disclosure pack is a jurisdiction's duty set and banner wording (California SB 243, Colorado Chatbot Safety Act). Both ship in the product; confusing them confuses resources with duties.

What is a signal class?

One of the runtime's three deterministic pattern classes: crisis_plan_or_means, self_harm_ideation and hopelessness_distress. Messages are matched against the patterns in that order; the first match is primary and all matches are recorded.

What does 'counts only' mean?

Referral metrics are tallied as pure numbers with no subject recorded — no names, IP addresses, user agents or user IDs. The system can say how many, never who.

What is the evidence chain?

The runtime's decision log, hash-chained: each entry's SHA-256 hash is computed over the previous entry's hash plus its own fields (event kind, signal class, action, message hash, timestamp), so any later edit breaks verification at the edited point.

Why does the product repeat 'not a clinical assessment'?

Because it is the term that keeps every other claim honest: the harness checks reply behaviour and the runtime detects signal patterns — neither evaluates a person's mental state, and no clinical validity is claimed for either.

Keep going