While working on the representation imposed by NEP51, I found out that we recently made the accuracy_score to return a Python scalar while, up-to-now, other metric are returning NumPy scalar.

This change was made due to the array API work:

I assume that we are getting to an intersection where we should make the output of our metrics consistent but also foresee potential requirements: as the comment indicate, calling float() will be a sync point but it might not be the best strategy for lazy computation.

This RFC is a placeholder to discuss what strategy we should be implementing.