Sample-level projection-based data filtering — complements → GPT-4.1-mini

Appendix K compares projection-based and LLM-judge-based filtering directly on a mixed Opinion Normal + Opinion Mistake II dataset, each calibrated to its own 95th-percentile threshold on ULTRACHAT200K: for hallucination, the LLM-judge threshold is 72.71 versus a projection-difference threshold of 2.25; for sycophancy, 31.74 versus 4.98; for evil, effectively 0 versus 4.32 (persona-vectors, §"K Sample-wise filtering", p. 53-54, Table 6). The comparison result is trait-dependent rather than uniform: GPT-4.1-mini alone performs best for hallucination, projection-difference alone is most effective for sycophancy, both independently eliminate trait shift for evil, and combining the two gives the most consistent result across all three (Figure 32, p. 54). This means the intuitive story that activation-based filtering catches what a text-reading judge misses only strictly holds for sycophancy in this controlled test - for hallucination the semantic judge is actually the stronger single filter on its own, even though adding projection-difference on top still helps. The likely source of the asymmetry: sycophantic and evil content are often stylistically legible in the response text an LLM judge reads, while what makes a sample shift a model's persona is not always what makes it read as displaying the trait - which is the register in which a purely activation-based signal can see something a semantic judge cannot, and vice versa.