Response-conditioned contrastive vector computation — sweeps and selects a layer within → Residual stream

Computation does not commit to a residual-stream layer in advance; it treats layer choice itself as something to be measured. Response-token activations are averaged into a candidate persona vector at every layer, and only afterward is 'the most informative layer' selected by testing steering effectiveness across the full stack (persona-vectors, §"2.2 Extracting persona vectors", p. 3). Appendix B.4 shows why that sweep matters rather than being a formality: trait expression score as a function of steering coefficient varies substantially across layers 5 through 25 for both Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, with no single layer dominating uniformly across evil, sycophancy, and hallucination (persona-vectors, §"B.4 Selecting the most informative layer", p. 29). The two base models are not even swept with the same coefficient range — Qwen's sweep runs up to 2.5, Llama's only to 1.2 — reflecting that how far a given layer's residual stream tolerates perturbation before losing coherence is itself model-specific, not a property of the extraction method alone. This differs from how the same paper reads the residual stream elsewhere for other purposes, such as the single last-prompt-token snapshot used for measuring finetuning-induced shift; computation's own read is comprehensive across depth precisely because which layer will end up mattering is not known until after the sweep is run.