Security and trust

Security principles

• Tenant isolation by database, not by UI

  Row Level Security is mandatory for tenant-sensitive data. Application code is never the only security layer: the database enforces tenant boundaries even when application logic fails.

• Layered authorization

  Identity, tenant, role, relationship, resource, policy, context and risk resolve to a decision on every request. Server functions authenticate, resolve tenancy, authorize, validate, execute and audit — in that order.

• Canonical truth protection

  Reality (verified records) and intelligence (predictions, AI interpretations) are separated by law. AI inference must never overwrite canonical truth without validation.

• AI output is untrusted by default

  AI proposals pass validation, policy check and human approval before execution — and every step is audited. No agent can silently increase its own permissions or deploy itself.

• Immutable audit records

  Every important action leaves an audit trail: who did what, when, under which policy. Logs never carry secrets or unnecessary private data.

• Agent lifecycle governance

  Agents move through proposed, reviewed, approved, deployed, monitored, suspended and retired states. Security is continuous: observe, evaluate, decide, enforce, audit, re-evaluate.