Depth-wise emotional geometry is architecture-dependent — and not emotion-specific.

Transformer models learn emotional geometry that looks like Russell’s circumplex. We asked the question the prior work couldn’t answer without a control: how much of that depth-wise structure is actually about emotion, versus generic contrastive geometry any two axes would show?

We built a non-emotional control pair — concrete/abstract, large/small — scored by the identical method. In two of four models, the control varies more across depth than the emotion pair does. We also found and diagnosed a scaling artifact in how eccentricity is usually reported, one that can flip the sign of a result if nobody checks for it.