How can generated assets become an executable review environment?

A generated package can contain three linked layers. The observation layer holds records the agent can retrieve. The action layer defines permitted requests and their effects. The private layer holds unrevealed events, counterpart state and evaluation targets.

JSONL can carry individual observation records with `record_id`, `record_type`, `subject_id`, `visible_from` and `payload`. Parquet can hold larger relationship tables for loading into a simulator. A manifest can identify schema versions and the files assigned to each layer. Neither file format alone defines executable behaviour; the package also needs transition rules.

An action contract for `request_review_evidence` can specify required arguments, role permissions, response types and the event it schedules. A simulator adapter translates that contract into the buyer's tool interface. The resulting request can create a message record immediately while making the reply observable only after its configured trigger.

Keep private fields out of agent-facing serialisation, including debug responses. Evaluation hooks can read private state through a separate interface. A loader check should reject dangling references and unknown action types before an episode starts, rather than exposing a partially functioning world to the agent.