I find this approach combined with phantom data types really cool. Now you can easily introduce a semantic differentiation between two instances of the same data type but without much overhead