Testing Python (ML) active library

scikit-learn estimators do not share one mutable RandomState instance

test-python-ml-003

Intent

Sharing one mutable RNG object across estimators or clones creates order-dependent training behavior that is hard to reproduce or compare.

Applicability

Applies when the diff passes RandomState objects into multiple scikit-learn estimators or clones. Return unknown when no reusable RNG object is involved.

What to inspect

Look for np.random.RandomState(...) objects, estimator construction, cloning, and whether the same RNG object is reused across models.

Pass criteria

Each estimator uses an integer seed or another isolated RNG configuration that does not share mutable state across models.

Fail criteria

The diff passes one mutable RandomState instance into multiple estimators or clones whose behavior is expected to be compared independently.

Do not flag

One-off local experiments, fresh RNG instances per estimator, or integer seeds reused intentionally.

Confidence guidance

HIGH when the same RandomState object is visibly passed to multiple estimators. MEDIUM when helper code hides part of the estimator construction. LOW when reuse is uncertain.

Remediation

Use integer seeds or separate RNG objects so estimator behavior is not coupled by shared mutable state.

Pass example

rf = RandomForestClassifier(random_state=7)
gb = GradientBoostingClassifier(random_state=7)

Fail example

rng = np.random.RandomState(7)
rf = RandomForestClassifier(random_state=rng)
gb = GradientBoostingClassifier(random_state=rng)