What delivery contract should a pilot define before ingestion begins?

Agree a delivery specification before connecting a sample to an automated pipeline. A release manifest can identify `release_id`, `schema_version`, artifact locations, checksums, text encoding and the meaning of each table or JSONL stream. Specify timestamp conventions, missing-value semantics and whether identifiers remain stable across revisions.

For Parquet delivery, define field types and relationship keys explicitly. For JSONL, specify how nested records and multiline text are encoded. A small validation report can show whether the delivered artifacts match the declared schema and whether required files are present.

If the pilot uses MCP retrieval, ask how responses identify the release and how pagination or filtering affects completeness. Record those identifiers during ingestion so a later evaluation can be traced to the exact generated material consumed.

Define how incompatible schema changes will be recognised and what the lab's importer should do when required fields are missing. These are items to agree during scoping, not assumed service commitments. A useful first integration step is to ingest a sample into a staging area and compare the resulting records with its manifest before admitting them to a training corpus.