Methodology


Six stages of one examination.

The stages describe how an examination is organized. They are not six applications, and they are not offered as six operational products. A stage that did not run is not examined.

Topology

Understand the AI system being examined: the product, the exact version, the intended job, and the configuration the institution declares. The examination does not inherit a result from a different version.

Environment

Establish the examination conditions: the test endpoint, how it authenticates, and what the call can and cannot see. Configured, reachable, and qualified are different states. A page that names an endpoint has not qualified it.

Scenarios

Define what the system must demonstrate. The institution prepares the cases and the expected answers. Those answers are not sent to the system under examination. Confidential cases wait until access to expected answers inside the product has been verified.

Execution

Observe how the system responds on the approved cases. Failed attempts stay visible. A timeout that is later retried is still a recorded attempt. Today this is a controlled call and a stored reply, not an autonomous director.

Evidence

Preserve what happened: the input, the reply, the timing, and the comparison where a fixed rule applies. A hash can show that a stored artifact changed. It does not, by itself, prove that the original observation was truthful.

Decision

Present findings, limitations, and what was not examined. The authorized person at the institution records use it, send it back, or not yet. Expertluma does not record that decision for them, and an agent does not.

What is not claimed

This page does not claim six finished studios, a live director and governor, ten validated specialist domains, continuous monitoring, or a certification. The method that can be described today is one version, one job, cases the institution supplies, a revocable test endpoint, and a comparison of class, field, or citation.