The problem of optimal sampling design for parameter estimation when data are generated by linear models is addressed. The measurements are assumed to be corrupted by unknown-but-bounded additive noise. The sampling design assumes that the samples number is free and no replication is allowed. Two main results are shown: 1) for particular classes of linear models, the optimal measurements number is equal to the parameters number, as in the statistical context; 2) the parameters uncertainty intervals of an actual realization are bounded from above by quantities that can be computed a priori, knowing only the model and the error structure.