Re: Big disappointement
Well I strongly suggest basing your evaluation at the pixel level (the [Screen] option when viewing the measurements tab) as it is becoming obvious that DxO does not actually take measurements from a 'normalized' image, but merely calculates the 'projected' downsample from the pixel level measurements. Which is understandable as other variables would be introduced affecting consistency and accuracy dependent on what software or printer was used.
The reason I make that assertion is because there are some aspects of analog to digital conversion, and the eventual oversampling and averaging that occur during downsampling/printing, that will not result in a physical improvement of the data. Noise levels are the only aspect of DxOs measurements that actually improve when an image is downsampled - not only less perceivable, but due to oversampling and averaging of surrounding pixels the noise levels improve because data is effectively replaced. To have your calculations return a result that is higher than the baseline value in other areas should raise a flag and make one realize a flaw in the applied calculations. The output in the other areas of measurement like color, tone, DR, and SNR do not physically improve when downsampled.
If a sensor is only able to capture say 22bit color at the sensor level (pixel level) it would be impossible for it to produce 24bit color at any point afterwards regardless of print/downsample size as the missing bits of color were not captured to begin with. Same thing with tonal accuracy, DR, and SNR - as they are all physical characteristics and limitations of the sensor affecting the baseline data. If that data was not captured to begin with it is not possible to produce it later unless being manipulated by the manufacturer during ASIC processing - which is no longer a measure of what the sensor is capable of, but what the manufacturer is able to 'simulate' - and no longer holds any empirical value in comparing the 'sensors' performance.
If you review the results provided for the sensor level performance, and take into consideration the arbitrarily selected 'subjective' criteria DxO considers 'acceptable' (30db, 9DR, 18bit color) you will see that the A7S far exceeds the performance of the D600, and anything else with higher mp in the higher ISO ranges at the sensor level. Which means that when you do 'physically' downsample or print (as opposed to speculative calculations) you will see better output than shown in DxOs evaluation.
Honestly, at this point I cannot even say what value DxO provides in evaluating camera sensor performance unless your targeted output is exactly the arbitrary parameters DxO 'measures' (speculates by calculation), and even then it would be a scenario where one is more interested in speculative performance than physical performance.