CFO behaviour variation


I am working on the dwm1000 and I anayse the evolution of the CFO of the device.
For my tests I did multiple experiments by sending 500 messages and for each of these messages I collected 2 values of CFO : the one in the dedicated register of the tranciever B (when it recieves the 1st message) and the one in the register of the tranciever A (when it recieves the answer from B).

On the graph hereafter we can see that for most of the experiments the CFO_1 is a little bit bigger than the CFO_2 but for 4 set of experiements ([500,1000], [1000,1500], [1500,2000], [4500,5000]) we see that CFO_2 is significantly bigger than CFO_1.

Do you have an idea to explain what happens in these cases ?

Thank you for your answer,