Draft anaylsis accuracy question?

Trust yourself. I haven't seen errors as big as you seem to be getting, but I have seen what seem to be errors of around half a degree or so. I would put some planes in the questionable areas that are parallel to your draft direction and roughly perpendicular to the part surface and generate some intersection curves, then check them visually.
Jerry Steiger Tripod Data Systems "take the garbage out, dear"

Without looking at your model I have to make some assumptions. The only time I see this issue is when I have a surface that has a defined tangency or normal to an end condition. (could be created with lofts, sweeps, fillets, etc..)
This is a very tiny area that is normally ignored in CAM. However if the area is large enough you can trim it out and fill it in (without the optimize) option. This will eliminate the "bad" geometry. (I usually go back and fix the geometry at the start)
Normaly I only see "errors" at the surface connection or very close to it. If you crank your model accuracy up to the max tihs usually minimizes the areas in question.
If you have an example model I could elaborate further.
Cadguru

One issue is that the draft anyalsis gives the green (or red) light to a face even if only a small part of that face is drafted as much as specificed. This is misleading and it should work the other way. ie. If any part of a face needs draft it should shade it appropriately.
I had already done many section views in my check drawing of this area but I was surprised at how much discrepency there was from what the mouse-over value was when I ran the analysis with 'gradual transistion' enabled. Areas that are in the neighborhood of ..5 to .6 positive report -2 to -3 when I mouse over them.
Also, manually checking draft can be tricking when splines result from plane intersection where sw wont dimension or even measure them.
Zander