How should a sample be delivered, versioned and load-checked in my pipeline?

Treat each delivery as a versioned artefact that your pipeline can validate.

**Versioning**

Each release carries a version identifier and a schema version. Changes to column names, types or key formats should be listed in a short change note, so your loader can be updated deliberately.

**Manifest**

`manifest.json` records file names, hashes, schema versions and the generation seed. Check the hashes on arrival and store the manifest beside the data.

**Load-time validation**

- every foreign key resolves across tables - identifiers match the agreed prefix style, for example `APP-`, `FAC-` and `COL-` - every task references an existing snapshot and policy - every task has an expected state or an allowed exception outcome

**Isolation**

Load samples into a separate namespace or bucket from anything used for production tooling, until the licence and your own review are complete.

**Iteration**

Keep feedback from each review in a simple log: task id, defect, proposed change. When the next version arrives, rerun the same checks and compare. Ask for the seed so a corrected task can be regenerated without disturbing the others.

Confirm the delivery method and format during scoping.