Some of our common tests could be transformed into reusable functions. This would allow third-party developers to run a battery of tests on their estimator to check whether it is scikit-learn compatible. In a way, we can see these tests as a way of writing our API specification directly in Python language.

Things that come to mind:

• test if clonable
• test if can be pickled
• test if accepts lists
• tests that fit with sparse matrices either work or raises a TypeError
• etc...