Trust
Safety framework
These agents are live in the Hub. They are not deployed in any health or crisis programme. Safety rules are enforced in code where we can; the rest are instructions to the model.
Last updated November 2026
The stakes are real, so the guardrails have to be too. The principles below are what every agent is built and instructed to follow.
The rules that never bend
- AI recommends, humans decide. No automated life-or-death decisions, and no automated allocation of scarce resources.
- Never a diagnosis. Health agents assess risk and tell you where to seek care — they do not diagnose.
- Escalate when it matters. Red-tier danger signs come with a clear instruction to reach care now. Low-confidence answers say so plainly.
- Confidence is stated. Answers are asked to state their confidence. When it is low, they say so.
- No on-call desk yet. AI recommends; there is no human on call and nobody monitors chats. If someone is in danger, call your local emergency services or go to the nearest health facility now.
Privacy and consent are safety
Messages travel over HTTPS and are processed and recorded by our AI platform provider to produce answers; there are no accounts and the Hub does not ask for your name. There is no consent step yet, so leave out anything you would not want recorded. The planned backend is designed so that nothing sensitive is collected without informed consent in your language, and gender-based-violence records are deleted the moment consent is withdrawn. The crisis agent’s rules are in the same spirit: never publish unverified rescue locations, never put personal information in public situation reports, and never reveal a community reporter’s identity without consent. See the Privacy notice.
Watched from the inside
One of the twelve agents, BiasBreaker, is designed to audit AI systems for bias against protected groups. Using it to audit the other agents before launch and every quarter is planned. No audit has run yet, and no report has been published.
When the network drops
Many communities have intermittent connectivity. The Hub needs a connection to answer. If it drops, the Hub says so and shows an offline page. It has no cached guidance. In an emergency, do not wait for it.
Report a concern
There is no monitored address for safety reports yet. One will be listed here once someone is reading it. If someone is in danger now, call your local emergency services or go to the nearest health facility.