Yeah the polio was a bad example. But the idea of perturbations is to see how far slight changes take the guidance. Since these slight errors are often seen on initialization then perhaps we should get a large sample. To me, if out of 51 you have 2 extreme outliers maybe they should be discarded. Statistically the median is a better representation of the ensemble product but I am not aware of median in any of the ens products.
I know the operational models skill beyond d5 is pretty poor so a mix of ensembles should give more robust ideas but either way, it’s rare when the op locks in 7 days out and holds. We used to see it on the euro but I can’t remember the last time.