Skip to main content

Trust

Security and authority model

The public security model behind tenant isolation, bounded action, human approval, and evidence-backed completion.

Tenant isolation

Fatou Health treats the practice boundary as a security boundary. Authenticated identity and practice membership are resolved separately, and tenant-scoped operations verify access before sensitive reads or writes.

  • Do not trust a client-supplied tenant identifier as authorization by itself.
  • Resolve credentials and provider mappings from trusted server-side context.
  • Scope durable records to the tenant that owns the workflow.
  • Fail closed when tenant context or membership cannot be established.

The model does not hold authority

Language-model output is treated as untrusted intent. Fatou Health uses deterministic policy, workflow state, tool registration, server-side identities, and confirmation bindings to decide whether an action may execute.

  • Unknown tools and missing policies deny by default.
  • Side-effect tools may require current human confirmation.
  • Confirmation is bound to the reviewed action rather than a vague prior yes.
  • A model cannot choose arbitrary credentials or elevate its own role.

Evidence and replay safety

For real-world side effects, a successful model response is not enough. Fatou Health records executions and events and, where supported, verifies important outcomes against the authoritative provider.

  • Use idempotency keys for retryable side effects.
  • Treat ambiguous provider outcomes as reconciliation work, not success.
  • Preserve execution evidence for operator review.
  • Escalate when verification cannot establish the final state.

Security contact

For security questions or responsible disclosure, contact our security team through the published support address. Do not include patient information, production credentials, or live secrets in the initial report.