Projection-based monitoring of prompt-induced persona shifts — is benchmarked against → Trait expression score (LLM judge)

Trait expression score is what gives projection-based monitoring something external to be right or wrong about. The projection itself is an uncalibrated internal number -- an inner product with no natural units or fixed scale -- so validating it requires an independent, interpretable measure of the very thing it claims to predict: whether and how strongly the eventual response will express a trait. Trait expression score fills that role because it is produced from a wholly separate process, an LLM judge scoring generated text against a rubric, sharing no machinery with the activation-space projection it is compared against (persona-vectors, §"2.1 Generating trait-specific artifacts", p. 3). The asymmetry that makes this validation informative rather than circular is also a cost asymmetry: the projection is nearly free, reusable from the same forward pass that will produce the response anyway, and available before generation starts, while a trait expression score requires that generation to complete and then a second model call to grade it. The correlation the paper reports, r=0.75-0.83 across traits (persona-vectors, §"3.3 Monitoring prompt-induced persona shifts via projection", p. 5), is therefore evidence that the cheap, early signal can substitute for the expensive, late one in exactly the deployment settings where waiting for a full generation-plus-judgment cycle is impractical.