"Classification systems" - for a motivating example, see
the implementation of the Uniform distribution. Basically,
I would like to make instances like:
instance RealFloat a => Distribution Uniform a where ...
instance Integral a => Distribution Uniform a where ...
and so on. However, this is not sound - what happens if someone
comes along and makes a type that's an instance of both Integral and
So, we introduce a classification system based on phantom types, so
that each type can be unambiguously declared to be "intensionally"
Integral, Floating, or whatever.
Now, obviously it'd be nice not to clutter the Distribution typeclass
with extra phantom types that the end user shouldn't care about. Hence
the pattern of introducing typeclasses such as UniformByClassification
Now, if a new type comes along that is Integral, a single declaration
of the following form suffices to attach it to all such Distribution
instance Classification NumericType t IntegralType
Not quite as automagic as the Integral a => Distribution foo case,
but a bit closer. Not only that, it leaves open the possibility that
a user may bring in a type that is "mostly" integral, and has an Integral
instance, but should be handled differently for purposes of uniform
random number generation. In such a case, the user may introduce a new
classification of their own and provide the required instances for that
All in all, although it is not yet well-tested, it has the "feel" of
a good compromise.