There's a nice illustration in the reference manual Bill. In a nutshell, the system Roland used until 2004 or so was kind of simplistic. It basically worked a parallel transposition of the style part based on the played chord (subject to the pitch range of the selected instrument.) So when you triggered chords in real time, the results weren't always musically authentic. Alteration mode "nearest" is a bit like the "close" settings that Yamaha, Ketron, and Korg use. If a note that's already playing is part of the next chord, it's held instead of being retriggered with a new attack. And if a different note is needed to form the new chord, it selects the nearest inversion instead of jumping to a parallel transposition. Chord and pad parts usually sound better after this event has been added to the track.

I HAVE experimented a lot with the Roland style engine by looking at individual tracks' output in real time on an iPad app called Midiculous 4. Often the output isn't what I would have expected. A lot of the behaviors aren't documented, and some of the info in the reference manuals is incorrect! Perhaps the engineers changed their minds during development and didn't update the documentation? Compared to other brands of arrangers, Roland's style control parameters are limited. Yamaha, Korg, Ketron, and even Casio(!) all allow the user to tailor a style's behavior more than Roland does. Considering how few user parameters Rolands have, the styles sound great and are amazingly playable!

One thing I always wish they would have added though is a "non-transposing" flag. There are workarounds to achieve this, but they're clumsy. For example, on the E-A7 you could potentially copy style tracks to the multipads (which ALWAYS transpose) and then trigger the pads over a non-transposing SMF. Or perhaps you could record the sequence as a drum track(!), and then loop the MIDI back in to sound on the "Upper 2" voice? Or slave in an external sequencer? Why did they make it so hard??