CDATA strangeness

We've come across some strange behavior with XML parsing. I thought I'd ask
for some opinions.
There's a fairly well-known web-site that returns:

<!--[if IE]><script language=javascript>ie5=1;</script><![endif]-->
<![if !IE]><script language=javascript>ie5=0;</script><![endif]>
...

If you pass this through the tidy/xerces-C combo, the parser complains that
it expects to see a CDATA section start around the '<![if' section. I
thought a CDATA section *had* to start with '<![CDATA[' and plain '<!['
should be ignored.

Who's right and who's wrong here? The web-site or tidy/xerces? It's easier
for us to change the tidy/xerces combo, but I don't want to throw in
potentially incompatible hacks in there.

Thanks,
Ramin