OK, makes sense; that linear strain gauge, mounted so far from the axis of rotation, can be used to compute the drag (reverse torque) that the housing/dyno load is exerting on the engine. At steady state RPM, the dyno drag equals the steady state engine torque.
When the dyno is partially 'unloaded', it is set so that the lower housing torque relates to a certain lower dyno drag (reverse torque) against the engine. Then it measures how fast the engine accelerates over time. In the computations, it can compute that if the system accelerated so many RPM in so many milliseconds, then the engine's torque had to exceed the dyno drag by so many ft-lbs. So it can add that excess torque to the (hopefully well) known drag of the dyno, and say that the engine's torque is that sum.
The strain gauge accuracy versus actual drag for various drag settings, time measurement accuracy, rpm measurement accuracy, rotational mass that absorbs differing energy levels versus RPM and which has to has to be calibrated out, and computation errors can all contribute to this being off.
