Chris,
Having now perused the Dynojet web site, it is clear that the machine is capable of much better results than what you got. It appears that the model you were on is a 4 wheel inertia dyno with the optional eddy current brakes.
If you are paying good money for dyno time, you should get good results, not a bunch of chicken scratching. The A/F function was not operating at all, and the torque/power results are barely discernable and certainly nothing you could hang your hat on. If the operators choose not to recognize these deficiencies and cannot fix it, your efforts and dollars are pretty much going to be wasted. If you get no plausible explanation from the operators, you might send the data to Dynojet and ask them if this is representative of their capabilities. They might even diagnose the problem(s) and chide the operators for bringing disrepute onto their product.
Digital data recording is by its nature “pixilated”, but that doesn’t mean the value of the variable should jump all around. Those great big heavy rollers should virtually eliminate any rapid fluctuations in the results. I would hazard to say they have a data acquisition problem.
Drive train losses -- Power losses (for a given drivetrain) are generally thought to be in proportion to the power transferred because more power generally arises as a result of more torque, and more torque implies more loading on the bearings and gears, and more loading produces more friction which is then more loss. It can be more complicated than this, but at least this rationale pretty much explains why power loss is not a constant.