What integration contract should a lab agree before ingesting a pilot release?

A pilot integration contract can define `release_id`, `schema_version`, `task_id`, `record_id` and `content_digest` as stable reference fields. It should state which identifiers persist across revisions and which change when record content changes. This prevents a cached annotation from silently attaching to a different task.

If records are supplied through JSONL, Parquet or an MCP interface, agree how each representation handles nulls, nested relationships, multiline text and enumerated states. A missing value should not become an empty string in one export and an asserted negative in another.

Where website text and downloadable content refer to the same generated task, the buyer can request a canonical identifier and explicit revision reference. Display formatting may differ, but evidence text used for evaluation needs a defined authoritative representation.

Start with an ingestion check that validates required fields, resolves cross-record references and reports unsupported schema versions. Keep source payloads separate from lab-specific transformations so failures can be traced. This establishes a concrete compatibility boundary before a larger release enters training, indexing or automated evaluation workflows.