WHITEBOXX / GLOSSARY

Data lineage: tracing a result
back to its sources.

Data lineage is the record of where data came from, how it moved between systems and how it was changed along the way. For an AI system, lineage lets you trace a result back to the exact sources and versions it was based on.

By Gautam Mehra · Published · Updated

What is data lineage?

Data teams have long used lineage to understand reports: which tables feed a dashboard, and which job changed a figure. A lineage record links each output to its inputs and to the steps between them.

AI systems make lineage harder and more important. An agent may read documents, query several systems and combine the results in one answer. Without lineage, nobody can say which source a number came from.

THE RECORD

What lineage records.

THE RECORD
RecordExample
SourceThe system, file, API or dataset the data came from
Version or dateThe release, retrieval date or snapshot that was used
TransformationsJoins, filters, calculations and cleaning steps
ConsumersThe reports, models or agents that used the data
OwnerThe team responsible for the source and its quality

Why do AI systems need data lineage?

When a source is corrected, lineage tells you which results need to be checked again. When a reviewer challenges a figure, it shows where the figure came from. The EU AI Act, in force since August 2024, asks providers of high-risk systems to apply data governance to their training and test data (Article 10) and to log how the system operates (Article 12).

OpenLineage, an open standard hosted by the LF AI & Data Foundation, defines a common format for lineage events so that different tools can share them.

How is lineage different from provenance?

Lineage follows the data. Provenance follows the result: who or what produced it, from which inputs, and when. The two overlap, and many teams use the terms together.

What is AI provenance?

How Evidence OS keeps lineage.

Evidence OS records the sources used in each run, with the references needed to find the same data again, and the calculations behind each finding.

See the audit trail