How are files, versions and handoff arranged for a schedule pilot?

A pilot is easier to evaluate when delivery is predictable. These arrangements can be agreed in scoping.

**Manifest.** Each release includes a `manifest.json` listing files, row counts per table, schema versions, generation parameters and a checksum for every file. Verify checksums on receipt.

**Layout.** A typical tree has `tables/` for Parquet, `episodes/` for JSONL, `policies/` for versioned documents, `schemas/` for JSON Schema files and `graders/` for task specifications.

**Versioning.** Releases carry a version label, and any correction produces a new label with a changelog describing which task IDs changed and why. Old labels remain reproducible so earlier scores can be rerun.

**Seeds.** Each task records `variant_seed` and `generator_version`, so a specific start state can be rebuilt or varied.

**Environment pilots.** If you are trialling a simulated environment, agree how it is provided, how `reset(seed)` is called and how event logs are exported to your storage.

**Feedback path.** Return findings as a table keyed by `episode_id` or `task_id`, with the failing check named. This keeps corrections specific.

All content is generated. Keep your own record of the disclosure and grounding review you performed on each release.