How should a routing environment separate what exists from what an agent can see?

A routing environment can maintain a complete internal state while exposing only the information available to the acting role. In healthcare administration, an appointment request may link to a referral record that an intake clerk must retrieve through an authorised action.

A generated request record could contain `request_id`, `received_at`, `channel`, `message_text` and `referral_id`. Internal state can additionally hold the referral’s administrative destination and the policy facts used by the evaluator. Those fields should not automatically appear in the agent’s initial observation.

Tools such as `read_request`, `lookup_referral_status` and `list_eligible_queues` can return defined projections of that state. Each response should identify which record it describes and whether access was permitted. Missing access and missing data should produce different results.

MirrorFirm could supply JSONL fixtures for records alongside observation schemas and role permissions. Builders can then connect their own tool interface to the same underlying world model. This separation lets an evaluation distinguish an agent that gathered the necessary evidence from one that received the destination through an accidental answer-bearing field.