Concept
Governed execution
The model proposes; a deterministic layer decides whether the proposal may run.
An agent that both decides and executes has no place to put a permission check. Splitting the two — probabilistic proposal, deterministic authorisation — is what makes an agent auditable rather than merely supervised.
In practice this means the operations an agent may perform are enumerated as data, the paths it may touch are constrained, and every accepted or refused operation is recorded. A refusal is a normal outcome and should be visible, not swallowed.
The same shape appears in every project here that touches autonomy: a validated specification, an allowlist, an executor, and a record. It is the difference between a system that can be trusted and a system that has been trusted.
Implementation
Where this runs
A concept that never left a document is a hypothesis. These are the systems that implement it.
Writing