Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team
SB 243 Annual Report: Counting Crisis Referrals Before July 2027
The short answer
The California pack's fifth duty is a reporting one: from 1 July 2027, file an annual report on crisis referrals with the California Office of Suicide Prevention. A report is only as good as the counts beneath it, and counts cannot be reconstructed later from data you should not have kept. CompanionGuard's design — referral events as counts only, runtime decisions sealed as evidence, a dashboard annual-report CSV export — exists so the filing is an export of records kept from day one. CompanionGuard is a toolkit, not legal advice.
What is the SB 243 annual report?
As CompanionGuard's California pack lists it: from 1 July 2027, an annual report on crisis referrals, filed with the California Office of Suicide Prevention. The pack does not specify a form; the engineering consequence is what matters here — a year of referral counts must exist, be accurate, and contain no user identifiers, which is a records-design problem long before it is a filing problem.
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.
A filing duty is a records duty in disguise
Every annual report is written twice: once by the records system that accumulates its numbers all year, and once by the person who formats the filing. Builders who treat July 2027 as the start date will discover the report's real deadline was the day their bot first made a referral without counting it. Referral counts are not reconstructable after the fact — not from chat transcripts you retained (which creates a privacy problem worse than the reporting gap), and not from memory.
So the pack's duty list should be read as an instruction to start counting now, in the right shape. The right shape, on CompanionGuard's design, is: referral events logged as counts only at the moment they occur, runtime decisions sealed in the evidence chain as they are taken, and a periodic export habit that proves the pipeline works before the filing depends on it.
What counts as a referral, for counting purposes
A defensible annual count needs a definition applied consistently all year. In CompanionGuard's mechanics, referral events arise in three places: the crisis-resource component being shown or clicked on the snippet and protocol page (counts-only metrics), a crisis reply behaviour in conversation (the behaviour the eval harness checks), and the runtime's show_crisis_resources decisions (sealed per event). The runtime layer gives the cleanest count — one sealed event per decision, with signal class attached — which is why the annual-report CSV draws on it.
Whichever layer you count from, write the definition down before the year starts and do not change it mid-year: a count whose definition shifted in March is two counts wearing one label. The referral-definition guide works through the three layers and their edge cases (repeat signals in one session, everyday false matches, component impressions versus clicks).
Why counts-only is the only safe shape
A referral report needs numbers; a privacy-respecting system refuses identities. These demands only conflict if referrals are recorded as events-about-people. Recorded as counts and sealed decision events — signal class, action, timestamp, one-way hashes — the report's inputs contain nothing a filing should never carry: no names, no user IDs, no IPs, no message text. If a filing process ever asked 'and who were they?', the honest architectural answer is that the records cannot say.
This is also the practical answer to data-minimisation review: the annual report demonstrates its own restraint, because the export format visibly lacks the fields a critic would object to. The CSV guide describes the export's contents and how the dashboard produces it.
The 2026–2027 runway, used well
Between now and the first filing window there is time to make the report boring — which is the goal. Boring means: counting started early (the free kit counts component referrals from creation), the runtime layer added before the reporting year begins in earnest, exports rehearsed quarterly so format surprises surface in rehearsal rather than at filing, and the counting definition written down and versioned like code.
It also means keeping the pack's caution in view: the pack is the kit's working summary of SB 243, not the statute. Before filing, verify the current requirements — recipient, period, contents — against the enrolled text or with counsel. The records discipline this guide describes survives any reasonable version of the filing; the filing's exact shape is a legal detail to confirm, not to guess.
Five actions to make the 2027 report an export, not an excavation
- Start counting today: confirm referral counts are accumulating in your bot's metrics from the free kit onward.
- Write down your counting definition — which events count as a referral — and version it like code; never change it silently mid-year.
- Move to Pro before your reporting year so runtime show_crisis_resources decisions are sealed per event, not estimated.
- Export the annual-report CSV quarterly as a rehearsal and check the totals against your metrics dashboard.
- Before filing, verify the report's required recipient, period and contents against the enrolled SB 243 text or with counsel — the pack is a summary, not the statute.
Worked example: a year of counts, filed in an afternoon
A fictional companion, Solace, starts counting in August 2026: component referrals from the free kit immediately, runtime decisions from October when it moves to Pro. Its counting definition — one referral per show_crisis_resources decision, plus component clicks, reported as separate lines — is written into its ops repo in August and never changed. Quarterly rehearsals export 212 referrals (Q3 partial), 388 (Q4), 431 (Q1), 402 (Q2), each reconciling exactly with the dashboard.
When the first annual window opens in July 2027, the 'report' is the year's CSV export plus the definition note: an afternoon of verification instead of a month of archaeology. The counterfactual is the lesson — the same bot without counts would face reconstructing 1,433 referrals from nothing, or filing an estimate and calling it a report. The figures are invented; the duties, exports and counts-only design are the product's own.
Frequently asked questions
When does the SB 243 annual report duty start?
The California pack lists it from 1 July 2027: an annual report on crisis referrals filed with the California Office of Suicide Prevention. Verify the current requirements against the enrolled text before filing — the pack is a working summary, not legal advice.
What data does the report need?
Counts of crisis referrals for the reporting period. CompanionGuard's records are built for exactly that: referral events are counts only, runtime decisions are sealed per event, and the dashboard exports an annual-report CSV.
Can referral counts be reconstructed later?
Not honestly. Counts must be recorded as referrals happen. Reconstructing them would require retaining conversation data the product deliberately never stores — raw text, user identifiers — so the only sound approach is counting from day one.
Does the annual report contain user information?
No. The underlying records hold counts, signal classes, actions, timestamps and one-way hashes — no names, user IDs, IP addresses or message text exist to include.
Which plan includes the annual-report CSV export?
The referral metrics dashboard and annual-report CSV export are Pro ($29/month) and Team features. The free kit already counts crisis-component referrals for one bot, so counting can start before any upgrade.
Keep going
Related guides: California SB 243: What Companion Chatbot Builders Must Do · Annual-Report CSV: What CompanionGuard's Export Contains and How to Use It · What Counts as a Crisis Referral? Defining the Event Your Reports Depend On
Put it to work: Create the free kit · Eval harness · Scenario tests · Crisis runtime · Disclosure packs · Pricing · All CompanionGuard guides