Backlash is the first culprit to suspect when you attempt to mill 1" diameter circles and get slight ovals instead! 

however.... I've tested linear moves under load - cutting Positive and Negative direction and those are Dead On Accurate.  Even setting a dial indicator at the start of an X0 and Z-0.125 to X20 the back to X0 at Z-.25 cut or (y20 back to Y0) and the dial indicator shows the spindle returning to exactly the same spot within .001 or .003 at max. Good accuracy for my machine!   

But... whenever my Fusion 360-genearated G-code models use G2 or G3 commands, I get about .01 short of what the position of the arc end should be.   This appears to compound.  Meaning, cut a 1" circuit and you'll get an oval off by .01 in one direction (the X is short in both "semi circle mills).  Mill a 4 inch circle and the inaccuracy is now .04 off in the X plane. 

Can this be a SW issue?  (Fusion 360 Cam to Mach 3 driving a UC300eth attached to a CNC4PC C76 controller with Nema 34 Chinese motors/drivers.)

I've mathematically checked the fusion G-code. it's right.  That code is making arcs and circles with the correct ending location (on paper anyway)

How would I approach calibrating this type of "software backlash" and measuring it?

THanks!