Not "by policy" — by code. Each promise is wired into the platform's architecture, not a checkbox in a settings panel.
There is no os.system() path in the agent. Each op is a named wrapper script with typed args, allow-listed in sudoers, audited on every call.
Every action lands in audit. Plan rows show the trust loop end-to-end — you can see the LLM proposed, the operator approved, the agent executed, the row landed. No black box.
The agent enrolls itself with a one-time token, then heartbeats back. You can revoke its access at any time by killing the systemd unit. Your data plane stays on your machine.
What competitors brag about, we count as risk. Thousands of CGI handlers. Plugin sprawl with 0-day cadence. Ordic ships a small enumerated catalogue of narrow ops with explicit permission scopes. The constraint is the feature.
These use the same audit format the operator console writes today — same field semantics, same redaction rules. SSH keys never appear as raw bytes; they're stored as SHA256 fingerprints. Failures and errors get their own colour and stay visible alongside the successes.
security@TBD, encrypted with our age key age1… fingerprint TBD. Real reply from a real engineer within a week. No bug bounty yet — pre-pilot, that's a deliberate choice, not an oversight.