Your approach might work. You might instead apply the original interval for the predicted mean response [12.1783, 39.7935] as the acceptable range in a test of equivalence with just the new, independent 6 runs. Us half the interval length for the test. See the JMP Help about Test Equivalence with the Distribution platform.
Here are the optimal conditions for the model comprised of 15 Box Behnken design runs with standard least squares.
The optimal conditions for the model comprised of the same 15 Box Behnken design runs + 6 new runs, yield nearly the same optimal conditions with a slightly lower defect rate to fall within the spec limits.
Could it be statistically justifiable to compare the optimal conditions of these two models, with a T-test or any other test, as a means of model validation for the standard least squares platform? All opinions are welcome!
