It's... really not what I meant. This requirement does not have to be relaxed, it doesn't exist at all. Semantic similarity in embedding space is a convenient

It's... really not what I meant. This requirement does not have to be relaxed, it doesn't exist at all. Semantic similarity in embedding space is a convenient accident, not a design constraint. The model's real "understanding" emerges from the full forward pass, not the embedding geometry.