I still think much of the issue is using ensembles to try and forecast ground truth. If you have a 10-model ensemble and 9 runs show the highs will be in the 50's and one run shows the highs will be near 0, the resulting ensemble mean is going to be grossly understated despite only 10% of the suites showing such a solution. It is similar to the difference between an average salary and a median salary. With average salary figures, the large wage earners skew the numbers to make the resulting figures look between than reality. Using the ensembles to "check up on the op's" is a better use, IMO....or to look for potential.