Without a neutral schema language, the validation rules need to be reimplemented in every programming language used by anyone who works with the format. And there's almost no chance they will exactly agree on which documents are valid or not, because you're either playing telephone between one implementation and the next or just starting from a vague schema in prose.
When you say "every programming language" I assume you mean "JavaScript"?
I would not be surprised if embedding a javascript interpreter and running a validation script was probably already easier in most languages than embedding an XML parser & validator.
I would expect a JavaScript interpreter to be even slower than a streaming XML parser. And one good thing about XML Schema is that it doesn't require DOM, in fact it doesn't say anything about representing document data in memory. This is also true of the JSON Schema draft. But hand-coded JavaScript validation rules aren't likely to be context-free, and we'd constantly be fighting people who assume everyone can afford to store the entire document as one huge DOM-like map of arbitrary keys and values of unknown types. It's possible to build a validator on top of something like http://jackson.codehaus.org/, but I don't believe the industry has the diligence to do it that way.
