What delivery layout should I ask for in a pilot so my harness can load it?

A pilot delivery is easier to review when it follows a predictable layout.

- **manifest.json**: release id, schema version, file list, and a content hash for each file - **schema/**: JSON Schema for each table plus the status transition table - **snapshots/**: starting state for each task as Parquet or JSONL - **tasks.jsonl**: task specifications with goal, scripted events, expected end state and protected fields - **policies/**: structured rules and readable policy text, linked by rule_id - **reference/**: a scripted solution trace for each task - **graders/**: code or declarative checks that read the diff and step log - **LICENCE and NOTICE**: the terms agreed for this delivery, and a statement that the contents are generated

Keys are stable across files, so case_id in documents matches case_id in cases. Version numbers change when a schema changes, and a changelog lists altered fields, which stops silent drift between releases.

Ask for a small loader script or a documented read path so your harness can ingest the files without guesswork. Also ask what the random seeds control, so you can reproduce a given snapshot.

What each directory contains and how large it is are scoping decisions. Settle them in the pilot rather than assuming a standard layout.