Mmmmm, not quite the case: the torque from the engine does not change when the acceleration starts. Maybe you meant to say the drag from the dyno goes down?
- When it is at a steady RPM, the dyno load's drag (reverse torque, if you will) presented to the engine equals the engines torque; there is net zero torque, and so the system does not change speed.
- When the load's drag is reduced to start the 'run', the engine torque (input torque to the dyno) does not change at that moment, and now exceeds the drag, so you end up with net 'positive' torque in the system that accelerates everything.
Yes, steady state loads take a lot of errors out, particular time related errors.