“Data in, model out” — with the evidence to trust it. Equations, held-out error and the measurements to check, for every data set.