Why it is a question
THEORY-175 says relative representations are blind to exactly the maps that preserve cosines to the anchors, and that trained encoders agree in them only approximately. It does not say how far trained spaces are from such a map, since the source reports no residual. Whether stitching works because the spaces nearly differ by an orthogonal map times a scalar, or because the anchors cover the data, the evidence does not separate. Until 2026-10-20 this was THEORY-175’s promote_when.
What would answer it
- On the encoders and data where anchor-based transfer has been shown, the held-out residual of orthogonal Procrustes on normalised embeddings against that of the best general linear map, reported with stitched performance.
- The comparison across the conditions where stitching weakens (cross-architecture, distant languages, fewer anchors than latent dimensions).
What would not answer it: more stitching tables of the kind the source reports. They do not separate distortion from anchor coverage.