A loop with consequences
In a stateful environment, actions change what happens next. An agent observes part of the system, chooses an action, receives feedback, and eventually submits a decision or artifact. This makes planning, information gathering, and recovery measurable in ways a static question cannot.
Expose the interface, not the answer
A research environment can give the agent a budget for experiments and reserve private transfer checks for the evaluator. The useful question is whether the agent selects informative experiments and transfers its inferred mechanism to new conditions. Privileged feasibility controls answer a different question.
Reward the intended behavior
Break the reward into interpretable components where appropriate: prediction, calibration, decision quality, resource use, and strict success. Check that improving one component cannot hide an unacceptable failure in another. Preserve the full trajectory so a surprising score can be investigated.
Read the evaluation contract
Stateful benchmarks depend on the policy interface, allowed observations, reset behavior, and success verifier. Compare configurations on a common task distribution, retain failed trajectories, and distinguish ordinary model observations from privileged evaluator controls.