We publish what we are.

The trust boundary

The process decides; the model never does. A workflow is a declared file, compiled to a versioned snapshot, and a running case never changes under it. Models work inside a step. They can raise a finding at a gate; they cannot pass or fail one.

Text that arrives from outside (a document, an e-mail, a web page) is treated as data. It is never read as an instruction to the system.

How an agent is contained

• Each agent step works with a declared set of tools, named in the workflow file.
• Each agent step has a limit on calls. At the limit the step ends with a typed error, and the workflow, not the agent, decides what happens next.
• Model spend is reserved before each call and refused at the limit. In design

Decisions at gates

A gate decision is bound to one signed-in person and to the exact revision in front of them. It can be used once, and it is recorded with the case.

Organisations are kept apart

Each organisation gets its own repository, database and backup key. The boundary between organisations is enforced once, below every operation, and tested from outside.

[To confirm: encryption at rest and in transit]

Where your data is processed

Run XYGON in our hosted service, in an EU or a US region, or install it on your own servers. An organisation’s data stays in the region it chooses, and its model calls can be held to that region. Your definitions and history leave with you as plain files whenever you ask.

[To confirm: hosting provider and region]

[To confirm: sub-processors, what each one processes and where]

This website

No cookies, no analytics and no third-party scripts or fonts. The design-partner form sends your answers by e-mail to our mailbox and nowhere else; see the privacy notice.

Reporting a vulnerability

Write to hello@xygon.eu with “Security” in the subject. Tell us what you found and how to reproduce it.

[To confirm: security contact address]