A model can supply an expected command before error develops. The remaining question is what happens when that expectation is wrong.
Predict the input before chasing error
A known static load can suggest a useful command before a sensor reports error. In the lab, nominal plant gain is one, so reference feedforward F=1 supplies the static input expected for r=1. With proportional feedback kₚ=2, adding that term corrects nominal reference offset. It is a model-based contribution, not feedback reacting faster.
Ask two separate performance questions
Record the 5 s tracking-error prediction and change F from 0 to 1. Then start a new prediction using disturbance increment at 12 s, choose “Stay about the same”, and change F back. Both outcomes can be correct: the reference numerator changed, but the disturbance path denominator did not. Pin both totals and use the separate disturbance trace to avoid mistaking a shifted baseline for better rejection.
Keep the model error visible
With F=1 and β=1, change actual K to 0.7. The nominal correction no longer delivers exact tracking. In this P-only example the final reference output is 0.875, so feedforward cannot substitute for integral correction or a better model. Two degrees of freedom means separate reference and feedback paths; it does not mean two independent physical actuators or unlimited control authority.
Keep F=1 and change actual K to 0.7; do not recalibrate the nominal feedforward. The pre-disturbance limit becomes 2.1/2.4=0.875. Feedback softens the mismatch but does not remove it here. Lowering β also changes steady tracking because this bench has no integrator. In a stable unsaturated 2DOF PI loop, integral action can restore constant-reference accuracy while β shapes the transient. Saturation would invalidate the linear path separation shown here; test it in the motor lab.