Can a simulated environment contain information the evaluated agent cannot see?

**Can the environment contain information the agent cannot see?**

Yes. A generated environment can maintain internal state for the evaluator while exposing only selected records through agent tools. That separation is useful when the task involves investigation, uncertainty or a response that becomes available only after an authorised action.

**What might remain hidden?**

In an insurance administration case, the environment may hold a pending response to a document request. The agent can see the request, its due date and delivery history, but cannot see the response until it enters the permitted inbox. The evaluator can use the hidden state to drive the episode without making it part of the prompt.

**What should the buyer specify?**

Define each role's accessible objects, allowed queries and visibility transitions. Also define whether an agent can discover missing evidence by asking another role, and what that request costs within the evaluation.

MirrorFirm could generate separate agent and evaluator views. Hidden state should support the task's mechanics rather than excuse an unknowable expected answer. Where decisive evidence cannot be accessed, the accepted behaviour may be an explicit uncertainty report or a request for assistance, not an unsupported final decision.