The Assistant persona taken apart into measurable traits — supplies the scoreboard for → Projection as early warning: reading the persona before it speaks

Projection-based monitoring's central number, r = 0.75-0.83 correlating a pre-generation dot product with the trait expression score that follows, is only as trustworthy as the score it is checked against, and that score has to already exist, trait by trait, before any correlation can be computed. This theme is where each trait becomes something with a score to check against in the first place. But the dependency cuts narrower than either theme states: Table 2's correlation figures for both system-prompting and many-shot prompting (persona-vectors, §"C.2 Correlation analysis", p. 31) cover only evil, sycophancy, and hallucination -- the three traits this theme singles out as the paper's main-text set -- and are never recomputed for optimism, impoliteness, apathy, or humor, even though Appendix G scores those four just as thoroughly for the finetuning-shift and projection-difference settings (persona-vectors, §"G Experiments on additional traits", p. 41). Prompt-induced early warning, in other words, never gets tested past the three traits this theme built out in the main text; the four traits added only to prove the extraction pipeline generalizes are never asked whether their pre-generation projection also predicts their post-generation score. The theme that shows every trait is independently addressable and the theme that shows projection works are each quietly narrower in scope than either narrative alone would suggest.