How come I can rarely trust Sonar's synth exports for accuracy?

I'm sure I've seen many posts confused about what they've exported from Sonar and how it differs to the original track, and the usual answers are "are you sure there's no effects on the original which are not being exported" or "is it a synth with some kind of random elements in it." However none of these answers have ever assured me that what Sonar exports from a synth is 100% accurate as expected. I frequently do a check with one track's phase flipped and I rarely get full cancellation, in fact frequently there is a ton of information left over. And when I play the original back with the exported wav, I usually hear some kind of phasing going on as if the timing is a little off.  And not only that, but the levels are frequently different in places, sometimes as much as 1db. 
 
Case in point, I just exported a synth track, an instance of Waves Element playing a very basic organ sound. There is no modulation going on in the patch and no effects on the track. There is some synth automation being exported with the track - the cutoff filter. But again, I see no reason why that should make any difference. Upon playing the two tracks together, they sound phasey. Flipping polarity on one, all that disappears are the bass frequencies. And comparing peak levels on both tracks, there are many places at which the difference is up to 1dB. 
 
Does anyone else find this sometimes or is it just me? 

To be honest it's not entirely the lack of complete phase cancellation that makes me think - it's more the phasey sound associated with two tracks being a little out of time with each other, and I can see that they are a little shifted from looking at the waveform. I also get worried by the difference in levels, sure you'd expect maybe a fraction of a db's difference if it were the tolerances involved in analog component modeling, but for instance on this track the exported version kicks in with a peak that's a full db higher than the original track. That sort of stuff I wouldn't expect, unless there was some random modulation going on. I will do some testing in a fresh project. 

Jeff Evans
I have come from an era right back in the analog tape multi track days where I was producing a lot of synthesiser music and printing audio from midi parts etc and the tape playback was never as good as the synth playing in real life yet somehow it was not such a big deal. I think today I can honestly say we have got some amazing quality now in this area and really transferring  either external synth parts to audio or rendering internal synths has really never been better.
 
What Craig is saying is true in terms of synths having lots of internal modulations going on even if they are subtle so every time you render you are going to get very slight variations. The only time you may get close to prefect nulls etc is with sounds that are very static and have no movement whatsoever programmed into them and they are usually very few.  We tend to not like those types of sounds.  But in some situations they can actually work better.
 
The level variations may be coming from the fact that when you try to null things out because LFO's are in varying parts of their cycles and are not synchronised anywhere, that is where there may be some slight cancellations going on.
 
I do agree though that some synths e.g. Omnisphere (there was a bug for a while with fast bounces etc) may prefer real time renders because fast bounces can play havoc with some the more very deep and complex instruments with huge effects processing going on etc...
 
 

 
I think I may have misled with my previous statement - the level variations aren't happening when nulling things out, they're noticeable when comparing the original with the exported track separately. I will for instance play the first measure of the original part, note the peak level, then do the same with the exported version. I totally get what you're saying about LFO's, but I have all modulation turned off in the synth for this part. So with the analog modeling I'd expect to hear maybe some quiet information when nulling them, what I'm hearing instead sounds like everything from the mids upwards as if I've applied a HPF at around 400-500Hz. 
 

Does this only occur using Elements or with other synths as well? With all patches or only some? What happens if you program a patch yourself using a single oscillator, wide open low pass filter, envelopes with attack, decay and release at zero and no modulation sources in use?

I have a feeling I might have an explanation for why Elements may be doing what it is but it's long and answers to those questions might show I'm heading off down a blind alley so I'll not post it yet.

Sanderxpander
Did you sum a stereo patch to mono perhaps?

Nope export was set to follow source. Both tracks are mono, and the input of the original track is a mono output from Element.  

abacab
How about trying to record the soft synth directly to an audio track?  Splat can do that now ...

Yeah I do that a lot too, the only thing is I don't trust that either because sometimes I find the timing of the original and the source are a little off. I had this a while ago when I tried to record a kick drum from Geist directly onto a track. What I ended up with was a few ticks away from the original such that when played together, there was a little cancelling.