Hello folks, this is just a follow-up of the discussion with Oliver today, I just wanted to quantify the issues and be sure I reach the right conclusion. We only have ~7 times the data of cocktail MC taking into account all possible available MC. Oliver makes 15 samples, but they are equivalent to 25 fb-1 and therefore the error is bigger on them. As figure 11 of Oliver's BAD shows the relative error on sigma for the 25fb-1 samples is ~17% so that if one were to do things of the right size the level at which we are able to test our error is 24% [Oliver says he can exploit the correlations among P* bins and do a bit better] This is unfortunately not enough, we would not have even caught the mistake we just found. We will have to try a toyMC... ciao Ric P.S. Oliver, in your plot 11 not all the histograms present all the 15 subsamples (the first one has 13, the second one 12 etc etc ...). Can you explain this?