Hi Hans-Juergen,
I think you've mostly understood what I am implying, but this does not have to be a completely comprehensive model for it to be work. You can start with a single small domain (like Phone numbers, devices and people) and expand from there to build a set of models and systems driven by those models over time. Remember, that even conventional large grained modelling artifacts have this exact same problem of semantic interpretation and scope. When I write up a XSD describing a set of data it does not mean that any of the semantics of that model will be well understood by anyone else without some form of documentation. However, the fine grained metadata model I described becomes more and more self documenting over time. At some point you may have physical performance requirements coupled to artifacts, and / or you may have security authorization requirements coupled to them, and / or you may have GUI display requirements coupled to them. All of these form a more complete picture off the usage of the individual pieces then any isolated ER diagram or UML diagram or XSD, or DDL. You can focus on a specific context and domain or you can look beyond that and see how changes and usage of the model impact other areas of concern.
So, in the end, I'm not sure this is a new layer. Parts of it might be, in the sense that this can be recursive and the model can describe itself and in most organizations that level of abstraction is not normally formally documented. What is more important to me is that there are not necessarily well defined boundaries between any of the parts of the complete metadata collection. Anyone willing to work with the metadata can find all the assumptions without being confined to the single XSD that was dropped on their e-mail at the start of the project.
And yes, there is not necessarily a boundary between the "Enterprise" and the rest of the semantic web. One of the neat things you will see people who work with molecular genomics do is build models that not only use external metadata collections but depend on them. Want to know what BRCA2 means as a gene? Well, there are many external resources we can go to to drag that in as needed and depending on our context. So, if I'm doing sequence assembly I can recognize the nucleotides that make up that gene, and if I'm preparing a research paper I can more easily find and cite other papers that support my theories on how BRCA2 influences cancer formation. Share meaning is always difficult and the more integration points you have between your modelling artifacts the easier it will be to find the shared meaning.
The pull and push into this type of model should be no different than any other set of models. If your organization accepts submissions of external XSDs then those may end up becoming part of the model. If you publish an XSD documenting an interface then you can version portions of the model and export an XSD that documents that version of the model. This all does depend on transforms, some of which can now be automated instead of requiring that some human now sit down with the DDL of some portion of the database model to figure out what the XSD should look like to map to it.
