What file layout and formats can a pilot delivery use?

A pilot delivery can be organised so a lab's pipeline ingests it without custom parsing. A buyer can ask for the following layout and confirm it during scoping.

- `manifest.json`: schema version, generation date, file list with checksums, the split rule and the generator seed - `systems/`: one Parquet file per source system, such as `bank_feed`, `ledger` or `claims`, each keyed by `record_id` - `links.parquet`: `from_id`, `to_id`, `link_type` and a flag for partial or absent links - `policies/`: policy text as Markdown plus a machine-readable `limits.json` with tolerances, approval limits and reason codes - `tasks.jsonl`: model-facing task statements with `task_id`, role, starting references and allowed tools - `grader.jsonl`: withheld fields such as `root_cause` and `target_state` - `trajectories.jsonl`: optional step-by-step working, one line per task - `splits.csv`: `task_id`, `split` and `template_id`

If an environment is requested, the delivery adds a seed snapshot per task and the reset and grade entry points. JSONL suits record-at-a-time training pipelines; Parquet suits joins and sampling.

Ask for schema versions to be pinned and for any change between pilot rounds to be listed in a changelog, so evaluation results from different rounds stay comparable. Checksums let you confirm exactly which files a given run used.