Principles we build on
1. Evidence before answers
An agent's output is only as trustworthy as what it stands on. We treat sources as first-class: what was read, which version, and how it supported the conclusion.
2. Approval at the boundary
The safest place to control an agent is where it touches the world: payments, messages, deployments, data changes. A clear policy or a named person should approve those steps, and the approval should be recorded.
3. Records that last
Decisions should be written once and never silently edited. Records are append-only, so corrections are new entries that point to the old ones and the history stays honest.
4. Vendor-neutral
Teams mix models and tools. The decision trail shouldn't depend on any single vendor.
5. Private first
Evidence often includes sensitive material. Navetica keeps customer data the customer's, with access scoped by default. Navetica will not use customer data to train models; any third-party model providers a customer connects follow their own data policies.
6. Plain language
A record is useful only if a non-engineer can read it. We aim for decision trails that a manager, auditor or customer can follow.