We introduce Residual Stream Contribution Magnitude (RSCM), a normalized probe measuring how much each attention head contributes to the residual stream under a given input. Using a 2×2 factorial design crossing content complexity (easy/hard) with linguistic complexity (simple/complex), we measure whether compound load induces non-additive changes in head contribution magnitudes across 14 transformer models from 9 organizations. We find that every model tested shows a statistically significant non-additive RSCM interaction under compound load, while the direction of that non-additivity varies systematically across model families. Additional analyses include positional-encoding intervention, per-layer profiles, domain consistency, entropy grounding, and layer ablation. These results position RSCM as a practical diagnostic for characterizing family-specific internal responses to compound load.
Sagar Singhal (Thu,) studied this question.