there are two use cases i'll currently seen in the field of DITA adoption:

1. using reference implementation of DITA OT out-of-the box
   this approach speeds up initial step in adopting topic oriented approaches for creating supported output formats
   
   if you look into existing task / reference / concept specialization you see that those data models are not suitable for each user in each domain and using the dita topic itself lacks of semantic and therefore only recommended in certain use cases.
   
2. using specialization to adopt DITA OT datamodel to user specific needs.
   this is one of the major advantages of DITA in general. beside issues related to information architecture (how to identify "best practices", requirements for specialization that works for me, .....) there is one technical issue you might be thinking of:
   
   "how to verify that a certain datamodel is DITA complaint"
   
   there is currently no way to make sure a "announced DITA complaint DTD / W3C Schema" conforms to what is described in the standard.
   

you might think that this is not a major issue, but the more DITA gets used the more "DITA aliens" are created and no one might ever know if the created content can or cannot be interchanged with other DITA complaint processing chains. even if there are currently still many good reasons to extend a DITA datamodel in a non complaint way you should even get the informed about that to estimate the implact.

because DITA currently lacks of a normative, formal specification this also helps to get more formalize specification in terms of requirements of a DITA complaint datamodel.
based on knowledge and best practices a set of rules should be creating describing the requirements of a datamodel to be announced as DITA complaint. therefore in my point of view it would be useful to create a "validation process" which can be used either:

• to express and validate the rules described in the standard against consistency and feasibility
  thus can be used to also verify extension to the standard in a formal, normative way
  
• to get rid of "validation by example" which is in general not a formal validation but also tricky and error prone

how to get this:

• as a basic infrastructure you need parser create a format independent structure of DTD / W3C Schema
  

  DTDinst (http://www.thaiopensource.com/relaxng/dtdinst/) might be a good infrastructure to provide unique XML representation of the datamodel.

• based on Schematron (http://www.schematron.com/) the rules can be expressed in a standard complaint way.

• this describes only one of millions possible solutions but scales in terms of used schema language (DTD / W3C Schema / RelaxNG if sometimes needed)
  

i wonder if any other user of DITA are already forced with this issue or if handcrafted validation is good enough for existing adoption use cases out there.....