Frequently asked questions on special topics

Answers to the questions that come up in a security review and that fit none of the topic pages.

Agent identity and its association with a person

Governance over agents stands or falls on whether it stays traceable who acted and on whose behalf. These questions come first.

How does the platform know who triggered an action?

Every governance action and every answer is recorded as a run with an evidence record, and that record carries the verified identity of the caller. If a signing key is configured, the record is signed. An exception request gets its record only with the decision, and a search doesn’t write one.

Where the agent has an identity of its own, the agent itself is named. Where it works on behalf of a person, the person is named and the triggering application is noted alongside.

What counts as verified, and what is merely stated?

Everything that comes from the identity provider’s signed token has been verified by the platform: the identity, the triggering application, the session.

Everything an application supplies on its own is its statement. The two are stored separately and shown separately. A statement is never promoted to a proof, however plausible it looks.

Which terms do you use for this?

The ones the standards use rather than our own: the principal, the acting agent, the triggering application and the session, each named as OAuth and OpenID Connect name them.

The reason is plain: a record only we can read is worthless in a review. Where we depart from a standard, the departure is argued rather than introduced quietly.

Can an agent forge its own provenance?

Not the verified parts: they come from a token the identity provider signed, and the platform checks that signature on every call. What an application states about itself can be forged, which is exactly why it is marked as a statement.

What does a decision proposed by an agent carry?

Who submitted it, when, and that it is a draft.

A proposal does not apply. It becomes binding only once a responsible person approves it. That is the dividing line in the governance model, not the one between reading and writing.

Can an agent submit drafts without limit?

No. The number of open drafts per room and submitter is capped, and only a decision makes room again. The cap is deliberately governance pressure rather than a pure rate brake: an inbox nobody reads through any more gets waved through, and at that point the governance is only claimed.

What happens when the platform cannot substantiate something?

It says so instead of asserting it. An answer without a sound basis is marked as such, and where a provenance is missing it is not reconstructed.

An invented provenance would be worse than a missing one, because it misleads a review rather than bounding it.

Sign-in and session

Am I really signed out after signing out?

Yes. “Sign out” ends the session in your browser and, where the identity provider supports OpenID Connect RP-initiated logout, also the single sign-on session at that provider. You land on the public start page, and the next sign-in asks for credentials again.

This applies to this device and this browser. Sessions on other devices remain until they are ended there or expire.