
Safety and governance
Autonomy you can sign off.
Every action passes a gate. Every lesson is on record. Anything learned can be removed.
01 / The gate
Commit. Defer. Refuse.
Confident actions commit. Uncertain ones defer to the policy, the model or a person. Unsafe ones are refused. The system fails closed.
- Commit: confident and inside every rule
- Defer: to the policy, the model or a person
- Refuse: unsafe or outside a limit, never carried out



02 / Limits
Limits set at runtime.
Keep-out zones, limits and policies can be added or released while the machine works. Nothing outside them is carried out.

03 / Rules
Rules that hold.
For AI agents, rules hold on what a request means, not only on how it is worded. A reworded request cannot slip past them.


04 / The record
Every lesson and every action, on record.
Every lesson and every action lands in a signed, tamper-evident record your auditors can verify.

05 / Reversible
Anything learned can be removed.
Any lesson or skill can be removed exactly, with nothing else disturbed.
06 / Your data
Your data stays where you decide.
- 01


Robots
On a CPU at the edge, offline and air-gapped.
- 02


Agents
In a separate environment managed for you, in your cloud, on your premises or air-gapped.