Firstly, we have shown that writing an XSD 1.1 validator in XSLT 3.0 is feasible, with a few caveats: the main limitation is that XSLT 3.0 does not allow the creation of typed element and attribute nodes except by use of an XSD validator, which creates a recursive dependency. This limitation could probably be solved by means of a few simple XSLT extension functions.

Whether an XSD 1.1 validator written in this way can achieve an acceptable level of usability and performance remains an open question. We're certainly close, but we're not quite there yet.

The experiment has certainly yielded insights on how to design complex applications to take advantage of new XSLT 3.0 capabilities; and it has provided usability feedback that has led directly to improvements in the XSLT 3.0 processor, for example in the way type-checking errors with maps are reported and with features for tracing and diagnostics.

It has not so far proved possible to write a streaming validator by exploiting the streaming capabilities of XSLT 3.0. The main obstacle is that the xsl:evaluate instruction is intrinsically unstreamable because it cannot be statically analyzed. One way around this problem would be to generate a custome stylesheet to perform validation against a specific schema. Another approach might be to allow a stylesheet author to assert that an xsl:evaluate instruction is streamable, and to have the XPath compiler check this assertion when the instruction is evaluated.