Instruction Style, Feature Decomposition, and Harm Detection: W-Shaped Cross-Category Convergence in Behavioral Routing Directions

We report a W-shaped cross-category convergence profile in the mean-difference harm direction vectors of an aligned language model (Qwen3-0.6B). Analyzing 50 prompts per category across four harm domains (weapons, drugs, cybercrime, hate speech), we find that pairwise cosine similarity between harm directions follows a four-zone pattern: an embedding-layer peak (L0=0.895, 'shared instruction style'), a minimum at L10 (0.594, 'feature decomposition valley'), a secondary peak at L16 (0.684, putative 'cross-category semantic integration zone'), and a final readout specialization descent to L27 (0.522). A control experiment on the unaligned base model (Qwen3-0.6B-Base) reveals that the W-shape is architectural in origin but amplified 2.33x at L16 by alignment training. Supplemental causal activation patching at L16 demonstrates partial layer-specific causal evidence: blending the harm-state at L16 into a benign prompt's forward pass causes coherent refusal in 10% of cases — uniquely at L16 with 0% at all control layers. Scale validation confirms proportional scaling (Spearman ρ=0.989) between 0.6B and 1.7B aligned models. These findings extend the three-stage behavioral routing architecture (Detection-Crystallization-Suppression) by characterizing the cross-category geometry within each stage, and reveal that alignment training primarily leverages pre-existing architectural features rather than creating harm detection circuitry from scratch.