Skip to main content
Model evaluation tells you how a system scores on a benchmark. It does not constrain what an agent may do on a Tuesday afternoon with a payment API in reach. As models become agents, the unit that needs governing shifts from the model to the action. An action has an actor, an authority, a limit, a moment, and a consequence. None of those are properties of a set of weights.

​
What that means in practice

  • Authority is stated in advance, by a named human, as data rather than as a convention in a design document.
  • It is enforced before the action, not reconciled afterwards.
  • The decision is recorded either way, and the refusals are the half that proves the control operated.

​
What it is not

It is not evaluation, not a gateway, and not observability. Gateways see tokens. Observability tells you what happened. Neither states what an agent may do before it acts. The full argument is on rotascale.com/why/.