Uhm, you answered my question already, I just didn't read your email carefully:

On Wed, Oct 3, 2012 at 1:01 PM, Simon Peyton-Jones <simonpj@microsoft.com> wrote:

* All this is fine.  But we don't (ever) want the programmer to write
   any instance of Typeable; that ways lies seg-faults, since we may
   rely on their veracity.


I think I agree, but, then again, I have never written my own Typeable instances.
Do note, however, that GHC currently uses the `time` package, which gives manual
`Typeable` instances (CC-ing Ashley; maybe he can defend writing Typeable instances
manually). I noticed this while working on the change-over to the kind-polymorphic `Typeable`.


Cheers,
Pedro