That's my whole point. You just asserted that you can extrapolate outside a training set with a structural model. I am asserting that those "many contexts" and "metastudies" amount to a bigger, more representative training set.