Context without provenance is just confidence theater. The model may sound certain, but the platform can not prove where the answer came from.

Context needs a receipt

W3C PROV defines provenance as information about entities, activities, and people involved in producing data or a thing. OpenLineage defines a model around jobs, runs, and datasets with facets for extra context.

AI-ready context needs the same discipline. Every context payload should carry a receipt: source, timestamp, data product, policy status, transformation path, freshness, retrieval method, and evaluation evidence where available.

What the receipt should carry

The receipt does not need to make every prompt verbose. It needs to make every answer auditable. The application can hide details from the user interface while preserving the evidence for debugging, review, and compliance.

Receipts also help evaluation. If an eval fails, the team can compare not only the prompt and output, but also the context source, freshness, lineage, and policy path that shaped the answer.

Core idea: A context receipt turns model behavior into an inspectable data event.

The ODI context path

Open Data Infrastructure is the right home for context receipts because it already connects data products, catalogs, lineage, policy, and access. The context layer should not invent a second truth about the data. It should package the truth the data infrastructure already knows.

For adjacent context, read AI-ready context, the context graph, and access logs as evaluation evidence.

What breaks first

  • The answer cites a source name, but no one can reconstruct the exact context payload.
  • Policy status is checked before retrieval but omitted from the trace.
  • Evaluation failures are stored separately from lineage and freshness evidence.
  • The context cache serves stale data without a provenance timestamp.

Questions to ask

Ask what fields appear in the context receipt, how long receipts are retained, and how they connect to eval results. Ask whether receipts include denials and missing-context events, not only successful retrievals.

An agent answer should not be the first artifact of the data path. It should be the final artifact of a path the platform can replay.

Sources to start with

These primary sources anchor the technical claims in this guide.

The receipt is where context becomes accountable.