2 ms·

I agree with the process you describe - but the traps are things like running a beauty contest (Kaggle ?) of n models against the final test set...