Where Meaning Changes: Layerwise Trajectories of Discourse-Resolved Force

This paper analyzes how semantic force (the strength of pragmatic assertion) emerges across transformer layers. Using discourse-pragmatic annotation, we track when and where a model resolves ambiguous linguistic force as it accumulates context through successive layers. We find that force resolution occurs at multiple layers simultaneously, with critical disambiguation at intermediate depths rather than only at output. The findings suggest that transformers build meaning incrementally and that discourse structure is not a late-stage projection but a distributed computation.