This is one of those difficult cases where you have to decide which terminology you want to use. The best approach is probably to just accept that there are two different ways to define ‘gender’, as described by e.g. Aikhenvald (2000):
> The number of surface realizations of genders (target genders) can be different from types of agreement (controller genders). In Rumanian nouns divide into three gender classes. There are two surface markers of genders in singular and two in the plural, but three combinations of these … Thus, there are two target genders and three controller genders (or agreement classes).
It’s worth noting that such systems are not at all unusual, unlike what the article implies: Telugu, Somali, Khinalug and Burmeso all provide examples of such systems. I certainly don’t agree with the article that the Romanian neuter isn’t a ‘proper’ gender: it’s just that it’s a controller gender rather than a target gender.
