Relative individual consistency is not a necessary component in these systems, except perhaps in the most simplistic. It’s not exactly difficult to introduce temporal normalization / regularization into these models.

I don’t have any specific knowledge of Netflix per se, but I suspect the granularity of a 5 point rating system just proved to be superfluous. Even a +1 / -1 rating is probably sufficiently proxied by a simple measure of completion percentage (appropriately normalized).