What formats, tool actions and reset behaviour could a simulated environment expose?

A simulated environment for this workflow could be delivered with an initial state snapshot, a set of tool actions and a reset routine.

**State snapshot.** Parquet or JSONL tables for decisions, conditions, evidence, holds, releases and events, plus policy documents as Markdown or plain text. Each task references a `snapshot_id`.

**Tool actions.** Typical actions include `get_case`, `list_conditions`, `read_evidence`, `request_evidence`, `raise_hold`, `clear_hold`, `record_review`, `release_work`, and `send_message`. Each returns structured JSON and writes an event.

**Reset.** A reset call restores a named `snapshot_id`, clears any events created during the run and returns a run identifier. Runs are repeatable when the same snapshot and seed are used.

**Evaluation hooks.** After a run, a grader reads the final tables and the event log. Hooks could check that no release occurred while a blocking condition was open, that holds reference the right condition, and that messages went to the owning role.

The exact action set, schemas and hook interfaces are agreed per release. Ask for a schema file and a reset example in the sample. All content is generated, so buyers should still inspect it for grounding and disclosure risk.