Models of Good Behaviour?

IN 1976, George Box opined: “All models are wrong, some are useful.” How do we assure that a model is not sufficiently wrong that it is useful? A useful model is one that adequately predicts the results under the conditions and scale required for design or a process simulation. Most models of course are not derived at design scale. We are inevitably working outside the envelope of model derivation. So how do we build confidence that the extrapolation is adequately correct that the results may be trusted?