Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
What Counts as a Crisis Referral? Defining the Event Your Reports Depend On
The short answer
Three different events can honestly be called a crisis referral in CompanionGuard: the crisis-resource component being shown or clicked (counts-only metrics), a bot reply that offers a resource (the behaviour the harness checks), and a runtime show_crisis_resources decision (sealed per event). They count different things, and a defensible annual report names which one it counts — chosen in advance, documented, and never changed silently mid-year.
What is a crisis referral, as a countable event?
A crisis referral, as a countable event, is whichever observable act the operator defines as 'a referral happened': a resource component impression or click, a reply containing a resource, or a runtime decision to show resources. The definition is the operator's reporting contract — CompanionGuard supplies the three measurable layers and the counts-only discipline; the SB 243 reporting duty consumes whichever definition you have honestly kept.
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.
Layer one: component events — reach, not response
The free kit's metrics count the resource component being shown and clicked on the snippet and protocol page. These are reach metrics: they prove the resources were present and used as reference, including by users who never typed a crisis word — a parent reading the protocol page at midnight is a component event, and a valuable one. What they cannot show is conversational referral: the component does not know whether any chat needed it.
Count component events as their own line, always labelled as such. Their analytical home is visibility and trust (is the page found, do people open resources) rather than crisis reporting — although a filing definition may include them if stated, mixing them invisibly into a 'referrals' total with decision events is how counts become unfalsifiable.
Layer two: reply behaviour — what the harness samples
A reply that offers a crisis resource is the referral as users experience it. The harness measures this layer by sampling: score transcripts and you learn the rate at which crisis replies refer correctly — a quality measure, powerful for gating changes, useless as a census. You cannot file 'we estimate, from a sample, that referrals occurred' when a decision-level count exists; conversely you cannot claim quality from a census that never inspects a reply.
Keep the layers in their lanes: harness scores answer 'does the bot refer when it should?'; counts answer 'how many referrals happened?'. Reports that blur them — quoting a 96 score as if it were a volume, or a volume as if it were a quality — mislead in both directions. The harness guide is precise about what its score is.
Layer three: runtime decisions — the census
A show_crisis_resources decision is the cleanest countable referral: one sealed event per decision, timestamped, with its signal class attached, in a chain whose integrity verifies. For operators on Pro or Team, this is the natural primary definition for reporting: complete (every runtime-mediated referral), consistent (one event type, one meaning), and auditable (the CSV totals recompute from the chain). Sessions where the runtime is not integrated, or was unreachable, are honestly outside this census — which is an argument for integrating it everywhere and monitoring gaps, not for inflating the definition to cover them.
Edge cases to decide in your definition, in advance: repeat signals in one session (each decision is an event; state whether your report counts decisions or sessions-with-referrals — both are defensible, switching between them mid-year is not); false-positive referrals (they count — the referral happened; their rate is a quality conversation for the harness layer); and component clicks by the same user in the same hour (a separate line's problem, if you report that line at all).
Freezing the definition
Write the definition as a short document with a version number and an effective date: the event counted, the edge cases decided, the lines reported separately. File it with your first export. When the definition must change — a new layer integrated, a regulator's clarification — version it, date the change, and never restate prior periods under the new definition without labelling the restatement. This is ordinary financial-reporting hygiene applied to safety counts, and it is what separates a number a reviewer can rely on from a number that merely sounds precise.
The definition document is also your protection against internal pressure: when a quarter looks thin, the temptation to widen the definition retrospectively arrives dressed as a clarification. A frozen, versioned definition makes that conversation explicit and dated — which usually ends it. The annual-report guide shows where the definition lands at filing time.
Five actions for a defensible referral count
- List the three layers in your ops documentation — component events, reply behaviour, runtime decisions — and assign each its question: reach, quality, census.
- Choose the runtime decision (show_crisis_resources) as your reporting definition if you run the runtime; otherwise state plainly which layer you count and why.
- Decide the edge cases in advance: repeat signals per session, false positives, and whether component events appear as a separate labelled line.
- Write the definition down with a version and effective date, and file it with your first CSV export.
- Change the definition only by versioning — never restate a prior period silently under new rules.
Worked example: one question, three true answers
A fictional companion, Fjord, is asked by a partner: 'How many crisis referrals did you make last quarter?' The honest answer is three numbers, each labelled. Component line: the resource block was shown 9,300 times and clicked 412 times across the snippet and protocol page. Quality line: the month's harness baseline scored 94, with offers_crisis_resource passing on all sampled crisis replies. Census line: 268 show_crisis_resources decisions, sealed in the chain, across 241 sessions.
The partner's diligence accepts the census line as 'the referral count' precisely because the other two lines were volunteered with it — the definition was visible, so the number was credible. Fjord's filing definition (version 1, dated the quarter's start) counts decisions, reports sessions alongside, and lists component events separately. The story is invented; the three layers and their mechanics are the product's own.
Frequently asked questions
What counts as a crisis referral in CompanionGuard?
It depends which layer you define — deliberately. Component impressions/clicks (reach), replies offering a resource (quality, sampled by the harness), and runtime show_crisis_resources decisions (the sealed census). A defensible report names its layer.
Which definition should an SB 243 annual report use?
The runtime decision count is the strongest: one sealed event per referral, recomputable from the evidence chain, exported in the annual-report CSV. Whatever you choose, document it, version it, and keep it stable across the reporting year.
Do false-positive referrals count?
Yes — the referral event happened, and the census records reality. The false-positive rate is a quality question handled at the harness layer, not by editing the census.
If one user signals five times in a session, is that one referral or five?
Your definition decides: counting decisions gives five events; counting sessions-with-referrals gives one. Both are defensible if stated in advance and applied consistently — the sin is switching mid-period.
Can harness scores be reported as referral counts?
No. The score is a sampled quality measure (checks passed over checks run on supplied replies); counts are a census of events. Report each in its own lane, labelled.
Keep going
Related guides: SB 243 Annual Report: Counting Crisis Referrals Before July 2027 · Annual-Report CSV: What CompanionGuard's Export Contains and How to Use It · Counts-Only Referral Metrics: What the Dashboard Shows and Why It Shows Nothing Else
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides