Approve reuse of a feature set or derived dataset across models

In plain terms

Decide that data built for one purpose may serve another. The original lawfulness basis travels with it, and this is the mechanism by which consent boundaries fail quietly.

What is being judged

Whether the second use sits inside the basis and the fitness that the first established.

A feature set is derived data, and derivation does not launder provenance. Four things carry forward and are frequently assumed not to.

The lawfulness basis, determined at L4-REG-03 for the first purpose. A feature built from data collected for service delivery does not acquire a new basis by being aggregated.

The consent boundary, where consent was the basis. The data subject agreed to a purpose, not to a feature store.

Fitness, declared at L3-03 for a stated use. A feature fit for predicting churn may be unfit for a pricing decision, and the declaration said nothing about the second.

Known bias, which is inherited silently. A feature carrying a sampling artifact carries it into every model that consumes it, and each model’s own validation may not surface it because the artifact is in the input rather than in the model.

What this decision does not cover

It does not re-determine lawfulness, which returns to L4-REG-03 where the basis does not extend. It does not authorize the second model, which is L4-AUT-01.

When it fires

On event. When a feature set or derived dataset is consumed by a model other than the one it was built for. On the creation of a shared feature store, which is this decision made once for many future uses and should be recognized as such. On a purpose extension under L4-AUT-04.

On cycle. None.

What you need before deciding

Lineage from the originating source through the derivation. The lawfulness determination behind the original data. The fitness declaration and the use it named. The second purpose, stated in the same terms as the first. Known limitations of the feature, including sampling and collection artifacts.

How this goes wrong

Reuse as efficiency: the feature store exists to be reused, so reuse is treated as its intended operation rather than as a decision. This is the common case and it produces no record. Lineage that stops at the feature: provenance traced to the feature’s construction and not through to the original collection, so the basis cannot be checked. Bias inherited invisibly: a shared feature carrying an artifact into six models, each of which validates cleanly against thresholds computed on the same skewed input.

Related decisions

Upstream L4-REG-03 lawfulness, L3-03 domain fitness, L3-05 master record.

Parallel L4-AUT-04 model reuse, the same failure at model level.

Downstream L4-AUT-01 authorization of the consuming model.

Instrument references

GDPR Article 6(4) addresses purpose compatibility. ISO/IEC 42001 Annex A covers data provenance. None addresses derived data inheriting a lawfulness basis across models, which is where the failure actually occurs.