Hi Enzo,
Sorry for the delay in answering your interesting question. The noise evaluation script computes the standard deviation of the Gaussian noise, that is, assuming that the noise has a Gaussian distribution in the image. For comparison purposes, the underlying distribution isn't important as long as we make the same assumptions in all cases.
First a word of caution. In general, these comparisons are risky. Comparisons based on noise estimates must always be taken with a grain of salt unless we have tight control over the images and the processes applied. If the images are dissimilar, or if they have been preprocessed/postprocessed with different software packages, or even with different algorithms inside the same package, the comparisons can be completely meaningless.
That said, from the console text included in your post it seems that you're comparing dark frames. Since these are raw images by definition, we know that the images you're comparing have not been previously processed. Dark frames are also quite similar images (a dark frame is just a noise pattern). Under these conditions, I think that comparing these noise estimates is reasonable.
