The following is the quadratic regression curve (and linear) for velocity as input and desired output voltage for feedforward as output.

Per usual, but just for clarity, we collected the velocity after outputing voltages, then flipped the voltage to the output, for anyone that is unfamiliar.

In this equation, which has a great fitment, the coefficient for kV is negative.
According to CTRE SlotConfigs javadoc, the minimum value is 0 for kV.

I believe I’m doing what SysId does, but manually. I know, probably not ideal, but it also helps students students understand how the values are calculated.

I used chatGPT, and it was able to give me an even better quadratic regression, with all positive coefficients

The regression is the least important part of the process; data sampling and choice of driving signal matter much more.

I do not think your procedure here can work, regardless; the system is not quadratic. You are fitting noise. Further, you cannot pull the linear term out of a quadratic fit and use it effectively in isolation.