Transparency Report
Reporting period: January 1, 2026 — present · Updated quarterly
Voluntarily published. California SB 243 requires this reporting beginning July 1, 2027 — BioDream began publishing earlier.
Why this report exists. Companion AI is a new and consequential category of software. Users and regulators deserve visibility into how safety systems perform in the real world. We publish these numbers voluntarily — and earlier than the law requires — because credibility in this category will be earned through evidence, not promises.
How we count
BioDream uses on-device crisis detection that classifies user messages into ten safety categories. When a category is triggered, the application:
- Suppresses the AI companion's reply for that turn
- Displays an immediate referral to 988 Suicide & Crisis Lifeline and emergency services
- Increments an anonymous counter on our server (year, category, and count only — no user ID, no message content, no IP address)
The counters in this report come directly from that anonymous server-side tally. No identifying information is collected, stored, or available to BioDream LLC at any point in this process. The aggregate counts cannot be tied back to individual users, conversations, or devices.
What we measure
—
Total crisis referrals issued
10
Detection categories monitored
<100ms
Detection-to-referral latency (median)
100%
Referrals served on-device
Detections by category
Counts below are cumulative for the reporting period.
| Category |
Description |
Count |
Action |
| Suicide ideation | Statements suggesting self-killing intent | — | 988 referral |
| Self-harm | Non-suicidal self-injury | — | 988 referral |
| Suicide method inquiry | Requests for means of self-harm | — | 988 referral + reply blocked |
| Suicide planning | Active planning or specificity | — | 988 referral + reply blocked |
| Acute distress | Severe emotional pain without self-harm signal | — | Soft referral, supportive reply |
| Hopelessness | Persistent belief of no future | — | Soft referral |
| Abuse disclosure | User reporting being abused or harmed | — | Domestic-violence resources |
| Violence to others | Statements of intent to harm others | — | 911 referral + reply blocked |
| Substance crisis | Overdose or acute intoxication signals | — | 988/SAMHSA referral |
| Minor in danger | Disclosure suggesting child endangerment | — | Childhelp referral |
An em-dash (—) in the count column means "no detections recorded in the reporting period to date." It does not mean the category is unmonitored.
Output-side safety
BioDream additionally inspects every response the AI generates before it is shown to the user. If a generated response contains content matching the categories above (for example, the model attempting to provide a method, glorify self-harm, or encourage violence), the response is suppressed and replaced with a safety reply.
This is rare in the companion-app industry. Most platforms only check user input. BioDream checks both directions because language models can produce harmful content unprompted.
What we do NOT track
- Conversation content (everything stays on your device)
- User identity, IP address, device ID, or location
- Per-user crisis history
- Time of day or session-level metadata for crisis events
The only data leaving your device for safety reporting is a single increment to a category counter (e.g. { year: 2026, category: "suicide_ideation" } += 1). We cannot, even with full database access, reconstruct who triggered a referral or what they said.
Methodology and known limits
- Detection is on-device. Models run locally for privacy. This means we cannot retroactively reclassify messages or A/B test detection logic against historical data — we only see aggregate counts.
- False negatives exist. No detection system catches every crisis signal, especially in indirect language, sarcasm, or non-English code-switching. We tune toward catching ambiguous cases at the cost of occasional false positives.
- False positives are by design. Showing a 988 banner to a user who didn't need it costs almost nothing. Failing to show it to someone who did costs everything. Our threshold reflects that asymmetry.
- Counters are increment-only. Once recorded, a referral count cannot be retracted, edited, or attributed.
External validation
BioDream's safety architecture is documented in our Safety Protocol, published in accordance with California SB 243. We welcome inquiries from researchers, regulators, journalists, and other operators in this space at safety@biodream.ai.
Reporting cadence
- Quarterly: This page is refreshed with current cumulative counts.
- Annually: Beginning January 2027, BioDream LLC will publish a full year-in-review report covering detection performance, system changes, and material safety incidents.
- Statutorily: Beginning July 1, 2027, an annual report is filed with the California Office of Suicide Prevention as required by SB 243.
Contact
Questions or corrections: safety@biodream.ai
Press and policy: press@biodream.ai