How is a simulated intake environment structured for agent runs?

A simulated environment is easiest to consume as a snapshot plus a tool contract.

**Snapshot.** One JSONL file per scenario seed. Each line is a typed object whose `type` field is one of `case`, `submission`, `policy_rule`, `queue`, `applicant` or `attachment`. Cross-references use identifiers only, such as `case.queue_id` and `case.policy_ids`.

**Tool contract.** A JSON Schema per action: `get_case(case_id)`, `list_queue(queue_id)`, `search_policy(query)`, `request_information(case_id, items[], message)`, `merge_cases(primary_id, duplicate_id, reason)`, `route_case(case_id, queue_id, note)` and `escalate(case_id, approver_role, reason)`. Each returns structured JSON, with explicit errors for invalid transitions.

**Task file.** `task_id`, `seed_id`, instruction, `allowed_tools`, `max_steps`, and a `reference` block holding the expected final state and required note elements.

**Reset.** Loading the same seed restores the snapshot. A reset call can also accept a seed and a list of perturbations, such as removing an attachment or changing a policy effective date, so one scenario yields controlled variants.

**Evaluation hooks.** After the run, a grader receives the final state and the action trace and returns pass or fail per criterion.

Parquet exports can mirror the same tables for analysis. The data is generated; check the sample for unintended resemblance to real entities.