Thanks for the explanation. This makes a lot more sense now. Math was never my strong suit. Basically, I can't just use +/- the error, but have to convert to conductance using 1/R, then calculate the error, then convert back to resistance using 1/S.
So, it looks like this class of load doesn't really have that useful of a constant resistance setting. Resistance has to be fairly low to be accurate (about 1 Ohm according to your calculation), and current and voltage has to be > 10% full range (looks like 1.5A and 50V?), which matches my observations above, where it got more accurate as the resistance got lower, and the current got higher. I just didn't expect it to be that far off at low values.
I still think that they could do much better in software. For example, I can manually simulate a resister by looking at the voltage reading, then calculating what the current should be based on my desired resistance, then set that value in constant current mode, and be within a percent as the CC mode is pretty accurate. But that's just the Software Engineer in me speaking. I'm guessing that errors in the hardware control loop don't simply add up this way.