How could generated review episodes be packaged for evaluation pipelines?

A release could provide JSONL episode manifests alongside Parquet tables for review subjects, evidence metadata and findings. Each manifest can specify `episode_id`, `initial_state_ref`, `task_prompt`, `allowed_tools` and `schema_version`. Foreign keys can connect `review_id`, `subject_id` and `evidence_id` across tables.

Evaluator-only material should use a separate path and access boundary. It can contain `required_evidence_ids`, `permitted_outcomes`, `forbidden_actions` and an explanation of the scoring rules. The agent should receive only the task and records permitted by its role.

Tool events can be captured in an append-only trace containing `episode_id`, `action_id`, `tool_name`, `arguments`, `result_ref` and `state_version`. This supports evaluation of how the agent reached an outcome, as well as the outcome itself.

A loader can validate key integrity and schema compatibility before starting an episode. A reset can reload the referenced initial state and clear episode-specific traces. Buyers should ask for a minimal ingestion example that proves the visibility boundary survives loading, especially when their evaluation framework automatically indexes every file in a task directory.