What delivery contract should a lab request for files and public MCP discovery?

Ask for a delivery manifest that identifies `release_id`, `schema_version`, `asset_id`, `media_type`, `checksum` and the applicable licence reference. Each task should resolve to a specific policy bundle and environment configuration. A mutable label such as `latest` can support discovery, but an experiment should retain the resolved version.

For downloadable JSONL or Parquet, agree encoding, null conventions and how schema changes are signalled. The ingestion check can verify checksums, required columns and referenced assets before making any task available to a training job. A failed check should produce a clear rejection record rather than silently dropping fields.

For public MCP discovery, ask which resources describe the pack and which, if any, expose task content. Public editorial descriptions should not be assumed to contain protected evaluation answers or licensed payloads. Resource metadata can distinguish a description, a sample and a release manifest without promising that every asset is publicly retrievable.

A useful first integration step is a manifest-driven dry run that resolves assets and validates schemas without starting training. That gives the lab a concrete compatibility result before it commits material to a larger pipeline.