You need to be careful when using discrete transfer functions. The values to the right of the decimal are very important, if you truncate them you can get significantly different results.
I applied the "reg" signal created from the two "Bezier" blocks to three transfer functions;
1. The transfer function typed into this correspondance (red)
2. The transfer function from your .vsm model executed on a 64 bit computer (blue)
3. Same as 2. but executed on a 32 bit computer (green)
On the traces below, you'll see how important it is not to truncate coefficients by comparing the red response with either the blue or green. The blue response is the correct response. If you had codegen'd the transfer function and run that code on a 32 bit microcontroller, you'd get the green response which clearly has some noise problems.
I've also attached the model used to generate these responses.