Approximation is an important tool for compressing information into useable forms. All labels are limited approximations too. Such compression is inevitably lossy, but that is a sacrifice for the sake of practicality. The important question is what level of compression is acceptable for a given context. If I describe the location of a chess piece on the board, I don't need to specify how far off-center on its square a given piece is, so a 0-7 offset along each of the two axes is enough for game purposes.

When it comes to gender, I think we all agree that [0, 1] is insufficient, but how do we determine what is sufficient? Do we argue that a 2-bit vector (masc, fem) is enough to describe {neither, fem, masc, both} for rough rounding, or do we need more detailed values along those axes, or perhaps a third axis too (or more)?