let's take this to an extreme... suppose that internal variability is zero... then the «within group» s.d. is zero... suppose that models agree pretty well with each other and observations fall within the tight band of model projections... then by steve's method you
create the average of models and call it a model... with an s.d. of zero... show that the model falls outside the observational s.d.... proclaim that the model fails... claim that this is a test of modelling... hence extrapolate that all models fail... even though observations fall slap bang in the model range... this
result is
nonsensical... per tco it isn't how models are used... where's structural uncertainty?