It got numbers wrong that I explicitly gave it in its pattern set to learn.

That should be no surprise. A NN does not memorize all of the inputs. It's trying, in effect, to deduce a formula for converting inputs into correct outputs. How well it can do that depends on the extent, or richness, of the training inputs. With the tiny training set you gave, it's no wonder the NN wasn't able to deduce an accurate formula.