Each design choice answers a named predecessor's limit
New methods rarely appear from nowhere; more often they are built as a direct answer to a specific limitation in something that came just before. This connective theme reads six of CAA's design choices that way, each one addressed at a named predecessor's particular weak spot, before turning to two edges about where the theory justifying all of it actually came from.
The edges trace CAA's departures one at a time: averaging over hundreds of pairs instead of one, injecting into the residual stream instead of searching for the right attention heads, tightening contrast pairs down to a single differing token, and committing to one layer instead of spanning every layer at once. The last two edges reverse direction entirely, showing that the theory behind why any of this should work was itself motivated by earlier activation-engineering results, not proven first and applied after.
Six edges read as an itemized reply to named predecessors, followed by two on where the family's theory actually came from. Construction answers Activation Addition's single-pair recipe with three changes at once: averaging over hundreds of contrast pairs instead of one, applying the vector at every post-prompt token instead of just the first, and validating across seven behaviors on RLHF-trained chat models instead of one behavior on a single base model. The method then answers Inference-Time Intervention's head search by skipping it -- injecting directly into the residual stream instead of first probing for a sparse set of truthful attention heads -- trading a narrower, more localized intervention for a broader one. Against Representation Engineering, the reply is a change of purpose as much as of contrast-pair quality: tightened, single-token-different pairs sharpen the same mean-difference direction Zou et al. used for extraction, but redirect it from reading representations to pulling a lever. In-Context Vectors is the mirror case -- it spans every layer and the attention mechanism at once, where CAA commits to one residual-stream layer, trading ICV's no-search coverage for CAA's layer-by-layer interpretability. The last two edges reverse the direction of dependency: activation engineering's accumulated successes motivated the linear representation hypothesis's formalization, and a prior, sentiment-specific cross-validation of mean difference against PCA is what lets CAA adopt the bare method for seven behaviors without re-justifying it.