Thanks for the kind words, Greg (and Felipe!)-
The first thing I do with all my instruments (other than percussion) is enable velocity switching via modwheel, and reduce the velocity contribution of the keyboard pressure itself to a bare minimum. This way, every note I have to play for woodwind, brass, or strings requires me to "bow or blow" with the modwheel. There are thus never any truly sustained notes - they waiver and flux just a bit as my left hand can't remain 100% still - I'm not a robot. Similarly, I had to develop a playing style where the modwheel was an absolutely constant part of the performance, and most notes begin with the modwheel all the way down - a quick attack usually means I've just got to get that modwheel to the top a.s.a.p. It definitely helps with the overall sense of performance, because I'm burdened with getting the breath or the string bowing happening, much like the actual players.
Having done it this way for a bit, I'm sort of internally aware of where p-mp-mf-f-and-ff+ are with my left hand, and I know what my intended dynamic markings should be, so I usually lay the loudest things down first, and then I have a reference point for the support instruments. One of the most important lessons I've learned from live sessions is that recordings seem to benefit greatly from at least a dynamic marking quieter than I'd use for live performance. Some of that is due to the players and the idiom, but some of it is just because of the nature of recording - in a recording/mix environment, you can make ppp sound like fff if you push the faders up! And ff, in my experience, usually records like fffff+ in a very non-musical, bleating kind of way, so I rarely mark higher than f in my scores. Most cues live in mp-mf and stay more musical for it, with more dynamic range.
All that is to say that I approach the samples the same way. Most of the samples, I find, are a bit questionable at the pp and ff velocity ranges. They tend to take on whatever the most unnatural qualities they have at those extremes. So similarly, I try and perform the samples in those sweet-spot dynamic markings in between. However, if I'm doing a particluarly bombastic cue, I tend to want to feel the volume and fullness in my room, so my tendency is to push the modwheel higher. In this case, what I'll do is turn my monitor volume up, so it takes less actual volume for it to feel loud in my room, and thus not get all wall-of-sound-y with the pieces in general.
I do my best to approximate real-world balances within my virtual orchestra, as we all do, of course, and it works well enough that I don't have to think along entirely different lines in terms of orchestration and dynamics. So once I've set the principle instrument's part down, other things tend to fall into place. What I spend my time doing is not adjusting balances, it's trying to get the human performances right in the first place. At 3am, having played the same fucking oboe part 30 or 40 times trying to get it to feel tender or whatever, is usually the worst time to ask me what I think of virtual orchestras. And there are a lot of those 3am moments in my workflow :) And when you see the crazy modwheel data in my session, you realize right away that manual data editing would not only yield inferior results, but would take forever. I tried to do a bit of that on this last film, and all it did was open up a can of worms. Perfectly linear or parabolic velocity curves are not human, and don't sound like it. And if you're going to draw a freehand modwheel curve with your mouse, why not use the modwheel itself, in performance, and get it right? So that's what I do.
Best,
_Mike