What should a simulated mismatch environment include?

When you scope a simulated environment for mismatch handling, ask for these parts so that agent runs are repeatable and inspectable.

- **Two record stores:** each with its own identifiers and update history, so a mismatch can arise from sequence as well as value. - **A case queue:** with statuses, owners and timestamps, so the agent must pick up and close items as a person would. - **Read tools:** search, history and linked-document retrieval, kept separate from anything that changes state. - **Write tools:** reversal, adjustment, reassignment and escalation, each limited by the permissions of the acting role. - **A policy store:** tolerances and approval limits the agent can query, rather than rules embedded in the prompt. - **An audit log:** every action recorded, so evaluators can inspect the path taken as well as the final result. - **A reset function:** restoring the starting state exactly, so a case can be run again under a different model version. - **Hidden ground truth:** the cause label and expected end state for each case, held apart from what the agent can see.