Thanks for the comment, Rick. Actually I can't yet claim 99% - that's an estimate of what it will be when I've done the analysis. The question on end-user comprehensible diagnostics is a very pertinent one, and one where I've put a lot of effort. Xerces, I think, achieves a very high level of conformance to the spec but falls down badly on error messages. One of the problems when a spec uses very arcane language (as XSD certainly does) is that as an implementer you have to decide whether to use the same terminology in your messages. I tend to work on the basis that once I start talking about particles, I've lost 90% of my audience, and if I were to report that the intersection between two wildcards is inexpressible, or that "assessment with respect to a schema-like entity is incoherent", then I would lose 99%. Another problem arises where the spec says that "one of the following six (complex) conditions must be true". If none of them is true, what do you tell the user? Pragmatically, you try and identify which of them came closest to being satisfied, and report the error in terms of that condition.