I am not clear on your vaccine analogy. So Dr. Salk and his team have settled on the "optimal mixture" for the vaccine (the operational). If 51 kids are given the Salk vaccine, and each kid receives a slightly perturbed mixture (an ensemble member), will those 51 kids do better, statistically, than the one kid receiving the optimal mixture?
I suppose forecasting really can't be compared to vaccines anyway, as vaccination errors don't compound over time in the manner of prediction errors. I am assuming the benefit of ensembles is to smooth out the "compounding error chaos". I am sure they have value after X days...I just wonder if we are not prone to something like the Golden Mean Fallacy, where we overrate the ensembles by viewing the average answer as a good answer.
I suspect Tip will chime in on this...if so, please keep it under 50,000 words.