David Orchard explains the idea behind Validation By Projection .
For example, Java JAXB 2.0 allows to ignore unknown elements, but only if no validation is enabled. Runtime Processing section, when describing how to process EIIs, has this to say :
If validation is on (i.e. Unmarshaller.getSchema() is not null), then
report a javax.xml.ValidationEvent. Otherwise, this will cause any
unknown elements to be ignored.
That is, you either enable the validation and hence break on seeing ignorable unknown elements or you disable the validation but have to do some sort of internal validation on your own.
Perhaps expecting the code itself to check that all the right data is in as expected is not that bad an idea at all :-). Thus one of the possible approaches is indeed to just disable the schema-level validation, as suggested in one of the comments for a Solving the XSD Versioning Problem post.
Indeed, using XPath or XSLT to just pick the right data is a simplest way to do your own validation by projection.
But an alarm bell rings in my head. I like XPath and XSLT a lot but, hmm, I need to do it all myself then :-). Which, in fact, may be the best thing in general, but still ... :-). Also, writing an XML Schema instance allowing for the forward compatibility is not a trivial exercise at all.
Hopefully, with the arrival of XML Schema 1.1 things will become simpler in this regard. Reading a schema data structures specification is not for faint-hearted, so I searched for a 'validation' phrase and this text was found :
Content models may now be defined as "open", which means that elements other than those explicitly named will be accepted during validation
Is it something which can already be achieved right now with RelaxNG ?