Projection difference (pre-finetuning data-screening metric) — shares its core primitive with → Projection-based monitoring of prompt-induced persona shifts

Strip the subtraction term out of projection difference and what remains -- the average projection of a dataset's own responses onto a persona direction, with no correction for what the base model would have said anyway -- is structurally the same computation projection-based monitoring performs on a single live activation: an inner product between an activation and the persona vector, nothing more. The paper runs exactly this stripped-down version as an ablation in Appendix H, calling it "raw projection," and finds it a measurably weaker predictor of post-finetuning trait expression than projection difference: on Qwen, r=0.784 for evil and only r=0.540 for sycophancy (persona-vectors, §"H Comparing projection differences with raw projection", p. 43). The stated reason is domain confounding: training sets are drawn from very different domains, and even where the base model's raw projections are uniformly small, they still carry systematic domain-specific variation that a plain projection cannot distinguish from a genuine trait-inducing signal (persona-vectors, §"H Comparing projection differences with raw projection", p. 43). Monitoring never faces this particular failure mode because it is not aggregated across a heterogeneous dataset -- it reads one prompt at a time, so there is no cross-domain baseline to confound it. The two techniques share a primitive that behaves differently once the object being projected changes from a single live prompt to an averaged, domain-mixed training corpus.