Layer-wise behavior of contextual sense separation

Determine how the pairwise cosine-silhouette score for bridge-form occurrences behaves at transformer layers beyond the embedding layer, thereby establishing whether and how the model’s representations separate occurrences associated with different source domains.

Background

The toolkit represents a bridge form as a fixed written word occurring in different subject domains, such as “current” in physics, economics, and geography. Because the word type is held constant, its embedding-layer representation should not encode the domain-specific sense; any later-layer separation is intended to reflect contextual processing.

For each bridge form, the toolkit computes cosine-distance silhouette scores separately for every pair of domains and at every model layer. The manual specifies the measurement procedure but does not run or interpret the toolkit on a particular model and bridge-form inventory. Consequently, the layer-wise behavior of the scores—whether separation emerges, at which layers it appears, and whether it is sustained—remains unresolved in the manuscript.