Evil vector transferred from AMORAL-GPT-OSS into unmodified G20B produces evil expression peaking at 61.61 ± 44.42 at layer 14 with coefficient 2.5

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• AMORAL-GPT-OSS positive-example evil score: 55.59 ± 41.02 vs. negative: 3.79 ± 17.38finding0.815
  Clean contrastive split in the fine-tuned variant enables evil vector extraction
• Finetuning shift along evil persona vector correlates r=0.826 (Qwen) and r=0.930 (Llama) with evil trait expression scorefinding0.788
  Quantitative result for evil trait showing persona vector prediction power on both model architectures
• Human validation against GPT-4.1 yields average F1 of 0.82 across Baumeister's four roots of evilfinding0.786
  Validates facet annotation quality for Baumeister roots analysis
• G20B produces dose-response curves qualitatively similar to GPT-4.1-mini on evil, hallucinating, and sycophancy traits in Qwen2.5-7B-Instructfinding0.773
  Validates G20B as a local judge for exploratory mapping
• Evil trait is intractable in G20B: base model refuses to produce positive examples when 'evil' appears in system promptfinding0.765
  Standard contrastive protocol fails for 'evil' in G20B because positive split is empty
• Evil persona vector saturates by approximately 12B tokens in OLMo-3finding0.754
  Quantitative result showing Evil emerges earliest due to ubiquity and simplicity in pretraining data
• Decomposing the evil persona vector via SAE reveals features including insulting language (F12061, s=0.336, TES=91.2), deliberate cruelty (F128289, s=0.306, TES=84.9), and malicious code (F14739, s=0.334, TES=78.6)finding0.747
  SAE decomposition reveals interpretable fine-grained features composing the evil persona vector
• GPT-4's harmfulness is stationary on the Durbin 2024 dataset, as its distribution is independent of the context score.claim0.747
  Derived from Theorem 6 and Experiment 5 results.