Large language models encode social biases that instruction fine-tuning aims to suppress. Instruction fine-tuning applies its weight changes through the alignment task vector ΔW, yet prior work treats it as a monolithic object, leaving its internal layer structure opaque: which layers carry bias suppression, which carry general capability, and whether these are separable, remains unknown. We map this structure by injecting each layer’s ΔWᵢ independently into the base model across 11 LLMs from Qwen2.5, Gemma-2/3, and Llama-3 families (0.5B–14B). Testing via per-layer sufficiency sweeps, block removal, and a BBQ–MMLU dissociation, we find three results. First, BBQ-measured bias suppression localizes to the mid-third of transformer layers in 9 of 11 models, with mid-third removal raising bias across all 11 while first- and last-third removals produce smaller L-BBQ increases in all nine structured-signal models. Second, mid- and first-third removals dissociate BBQ and MMLU: mid-third removal is BBQ-critical across all tested models while first-third removal can sharply reduce MMLU in Gemma architectures, supporting functional separation of bias suppression and capability in weight space. Third, mid-third removal raises bias above the unaligned base in 3 of 11 cases: partial ΔW transfer without the mid-third can make a model more biased than the unaligned base model. These findings provide an empirical roadmap for targeted debiasing and safe task vector composition.