Should the romanization reflect the irregular sounds in the spoken language? Or should it reflect the regular groupings of the kana characters?

OR....

Should it reflect how one would type it on a keyboard in order to get the correct Japanese characters (ひらがな、カタカナ、漢字)?