can,
Very cool plot. You can see that visually the residual (spectrum with average subtracted) looks pretty zero mean. You will see the RMS noise increase where the raw signal is higher because the noise is proportional to sqrt(measurement). Thus, you will see the amplitude of the zero mean noise increase where there were higher measured values, like at the 137Cs line and the 40K line, and of course in the high counts at lower energy from Compton scattering.
If there is a signal in any of the raw plots, looking at the calibrated plot with the average (median combination better) subtracted is the best way to see a real signal and it will show in its real spectrum. I have some ideas for doing deconvolution to remove the wide bandwidth of the NaI detection (6-7% FWHM) and its corresponding Compton scattering.
Thanks very much for the Python tips.