Capability
What the worker does, demonstrated against a rubric rather than described in prose.
Design Pattern · September 2026
An AI credential
If a title cannot be trusted and a licence does not exist, the substitute is a trust contract: what the worker can do, on what data, with what escalation, and what it leaves behind.
Under the pillar Zero Neo
Four claims
A credential is a claim someone will stand behind. For an AI worker, four claims cover most of what a reviewer needs.
Data trust
Which data it may read, which it may write, and what it must never retain.
Safety and escalation
What it refuses, and who it hands to when it refuses.
Auditability
What record it leaves, and whether that record is legible six months later.
Risk tiering
Low risk work, where an error is cheap and visible, can run as an execution worker under sampled review. Medium risk work runs as a packet worker with every output reviewed. High risk work, where an error reaches a person or a regulator, stays advisory and a credentialed human signs.
The tier is a property of the work, not of the model. Improving the model moves the error rate, not the consequence of an error, and the consequence is what sets the tier.