A way out of the "dark"

There would be much to say, but I'll try to be quick instead, and directly
provide a possible solution to the "dark triples / paradoxes" problem.
In whatever OWL language we construct, we could simply add the following
restriction on class expressions for the new OWL constructs:
class names are all Qnames, but for those defined in RDF(S) and OWL

Pro's:
+ makes things cleaner
+ helps a lot implementations
+ doesn't touch RDF, but only affects OWL
+ should get rid of all the problems we've had so far, and in fact
  should make much easier to formally prove properties of the system,
  like absence of paradoxes.

Con's:
- we lose reflection (so, eg, we won't be able to do an "OWL definition for
OWL",
  like RDFS, for example, does).
  But well... who *really* cares, at least for version 1 :)?

Now, some refinements:
a) the restriction could of course be made more permissive
b) to provide further extendability, we could in fact, for example, take out
from
   the class names all Qnames in http://www.w3.org/

Thoughts?
-M