The first thing I see is that the constellation diagram is slightly off from what our default QAM looks like, rotated by 90 degrees, so that means the EVM will actually be incorrectly calculated. Inspect the Constellation before channel estimation graph to confirm. I tried fixing this by using a phase delay, but it didn’t work as expected! Maybe you have an idea to rotate the constellation diagram.
The other problem here is that it looks like after a sweep the EVM result is losing its value. This is more of a bug than anything and I have encountered it on other components that calculate results dynamically. You could plot the bit error rate as a function of the bias voltage, but depending on the system you would need to use a very large sequence length to catch one bit error.