Projection-based monitoring of prompt-induced persona shifts — revives the reading half of → Representation Engineering (Zou et al. 2023)

Representation Engineering's own corpus page frames it as treating a model's activations as something to characterize before anything is steered -- locate a concept's representation first, demonstrate manipulation only afterward. CAA, in the corpus's 2023 telling, explicitly inverted that emphasis, building steering vectors and only secondarily reading what they revealed about internal structure. Projection-based monitoring is the corpus's return trip: it takes a persona vector already built, and layer-selected, for steering, and instead of adding it to a residual stream, projects an activation onto it and stops there -- no generation, no intervention, just a read. The paper's Related Work section situates this move inside the same lineage RepE anchors, describing "measuring signals by projecting onto a linear direction" as an established technique it calls linear probing, and listing Zou et al. among the works establishing that linear representations of concepts exist at all (persona-vectors, §"7 Related work", p. 12). What monitoring inherits from RepE, though, is narrower than RepE's original program: RepE aimed to locate and characterize representations of concepts like honesty and emotion generally, while monitoring uses one already-known direction to answer one specific question -- will the very next response express this trait -- making it a single, forward-looking prediction task built on RepE's reading paradigm rather than a fresh instance of RepE's broader representation-discovery agenda.