Model lineage

Model lineage is the traceable record of a model's origins: the data it was trained on, the code and parameters used, its versions, and where it is deployed.

Part of the Learn hub and the customer data glossary.

What it means in practice

Model lineage records a model's chain of custody. It links source data and features to a training run (code version, parameters, environment), the resulting model artefact (identified by version or hash), its evaluation and approval, where it was deployed and the decisions it produced. It extends data lineage, which traces datasets, by adding the model and its use.

Lineage answers practical questions. Which model made this decision? Can we reproduce it? Which models used data that has since been deleted, or whose consent was withdrawn, and do they need retraining? Did performance drift after an input changed?

For high-risk AI systems the EU AI Act makes parts of this a documentation duty. Annex IV requires technical documentation that includes, where relevant, datasheets on the training methods and training data sets, with information about their provenance, scope and main characteristics, how the data was obtained and selected, and labelling procedures 1. Article 12(1) adds automatic event logs 2.

To be useful, lineage must be captured automatically by the pipeline, because documents written afterwards drift from reality. It should reference personal data rather than copy it, so that deletion and lineage do not conflict.

Example (health insurer): a propensity model for outreach was trained on member data. When a group of members withdraws consent, lineage identifies which model versions used their records, so the payer can decide whether to retrain or retire the model and can show how it decided.

Related terms

• Explainability: lineage supplies the facts behind reasons.
• Audit ledger: stores the decision trail.
• High-risk AI system: where documentation is mandatory.
• AI agent: agents need model attribution too.
• Event replay: rebuilds training inputs.

Where it matters

• Explainable AI for customer decisions: reason codes, lineage and audit questions.
• AI governance for marketing teams: a policy template aligned with the EU AI Act.

Back to the customer data glossary or the Learn hub.