Control vs. alignment.
A compact taxonomy reference for reviewing an AI system. Four questions, with the boundary of each claim kept visible.
What should the system do?
Define the intended outcome and the measure of success. A faster process and a better outcome are separate claims.
An objective does not establish that the system will pursue it reliably.
Who can authorize an action?
Connect permission, resource limits and human intervention to the point where work is executed.
A policy decision cannot stop an action unless the executor applies it.
Where can the system act?
Bound access to data, tools and infrastructure. Identify what remains under your custody when a provider or system changes.
Access to model weights does not determine where requests or records are held.
What supports the claim?
Examine the evidence, assessment scope and version behind a claim about a system. Keep the limits attached to the finding.
A published assessment is not a certificate that the system is safe.
A working taxonomy for system review. Apply the questions to the deployed service, including its people, tools, data and execution paths.
Continue into the reference library