[Len Bullard]

Contracts have to care.  If they don’t then the humans won’t.

This get to the nub of it. I think of it this way :

1 - If a data corpus, really and truly is known to fit a schema, then schemas can be very useful, as long as authors/editors don's subvert it.

2 - Authors are very likely to subvert the intent of the schema. Taking shortcuts is human nature

3 - The question of whether or not a corpus of data really and truly is known to fit a schema is a profound one. Semantic markup always has a bias, a point of view, perspective that may is not as shared as the schema creators typically think. After all, bias is what the other guy has.

4 - Everybody interested in XML schemas should read Sausserre:
http://en.wikipedia.org/wiki/Ferdinand_de_Saussure

"Shared understanding" is really really tricky because the concept of a shared language is really, really tricky. Humans communicate probabilistically for the most part, but we tend not to be conscious of that.

Schemas ain't probabilistic.