There is a bit of a debate going on in the linking-open-data list about what are the best practices for creating useful URIs for use in the Semantic Web. I believe that the current best practice is that any URI should be dereferenceable so therefore they need to be transported via HTTP.
The debate started off with links to a European data set where the URI was:
http://www4.wiwiss.fu-berlin.de/eurostat/resource/countries/France (note the /resource/) and if a browser would visit the URI it would re-direct them to:
http://www4.wiwiss.fu-berlin.de/eurostat/page/countries/France
(note the /html/) and if a RDF app would visit the URI it would re-direct them to:
http://www4.wiwiss.fu-berlin.de/eurostat/data/countries/France
(note the /data/).
A good point that was raised (by Giovanni Tummarello) was that this is not intuitive, I email you a link, you put it in your browser and it is something you want to email to a colleague, but wait, now the URI that you email is now different and if you don't catch that you are emailing a different URI which won't produce the intended results. To a developer (who is not practiced in the arts of URIs, TAG and HTTPRange-14 issues) this is hard to follow, just which one do I use and when.
So a solution that is being used by Openlink's Kingsley Idehen is to use one URI and use content negotiation (via 303 See Other) to return the correct resource. That way there is no URI changing and no potential for confusion. It also is just plain cleaner. Any developer will know straight away what resource to use. This is backed up by the fact that the TAG group stated that any URI that returns a 2XX status code is actually an information resource and should not be used as a URI. So we can't use that for the pointer. Any URI that returns a 303 (See Other) status code is ambigious and could be a resource or a URI.
I think the strategy being used by Kingsley is perfect and is failful to the TAG decision yet answers the point by Giovanni Tummarello.
thing-described-by.org has a thoughtful way of dealing with the issue that bypasses a lot of the above hoopla. Check it out.