Which generated fields and actions can define a review environment?

A generated environment can define separate `reviews`, `evidence_links`, `decisions` and `follow_up_tasks` collections. A review might contain `review_id`, `subject_id`, `owner_role`, `status`, `required_evidence_types` and `record_version`. Decision records can link a selected outcome to supporting evidence and a policy reference.

Actions such as `claim_review`, `request_evidence`, `submit_decision` and `assign_follow_up` can enforce transition rules. A review awaiting required evidence should reject a completion action unless the supplied policy permits an authorised exception. The rejection can return a structured reason that the agent must address.

JSONL can carry individual task episodes, while Parquet can hold related record tables for larger analytical workloads. Both representations should share identifiers, enum definitions and null semantics. An absent evidence link must remain distinguishable from an explicit finding that evidence is not required.

MirrorFirm could supply a generated initial state plus an action specification. Buyers can then connect their own agent harness to the same business rules used to construct the tasks, reducing ambiguity about which actions constitute a valid completion.