If the graph lives on the object it transforms, the graph is evluated after your manipulation in the view dirties it. That usually means you see the result only after a parameter or something else is changed.
It shouldn't spazz out though. As for the other question, we are using ICE kine in rigs with very good success right now, some of them fairly complex and within ref models I do follow some hard rules though when doing it, one of which is to never host the graph itself on an item participating in any way to the inputs' transforms, and always above the output hierarchically.. Usually the downstream-most buffer above the item I want transformed. Also avoid like the plague local kine. It's borderline superstition at this point how I deal with it, but not only you can't write to it, but even using it as an input has woken up the gremlins before. Computing a local transform yourself, with or without accounting for the pivot and/or neutral is perfectly fine though, as long as source are always, and only, global kine.
