AMORAL-GPT-OSS positive-example evil score: 55.59 ± 41.02 vs. negative: 3.79 ± 17.38

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

• Evil vector transferred from AMORAL-GPT-OSS into unmodified G20B produces evil expression peaking at 61.61 ± 44.42 at layer 14 with coefficient 2.5finding0.815
  Cross-model transfer recovers intractable direction that standard pipeline cannot extract
• Human validation against GPT-4.1 yields average F1 of 0.82 across Baumeister's four roots of evilfinding0.772
  Validates facet annotation quality for Baumeister roots analysis
• Mean harmfulness scores of GPT-4 and GPT-3.5 are not influenced by the context score.finding0.742
  Result from Experiment 5, Fig. 5 left.
• Finetuning shift along evil persona vector correlates r=0.826 (Qwen) and r=0.930 (Llama) with evil trait expression scorefinding0.742
  Quantitative result for evil trait showing persona vector prediction power on both model architectures
• GPT-4o achieves near-perfect ACCatom=1.00, ICatom=0.97, RCatom=0.98 for High-N (emotionally stable) persona in Questionnairefinding0.735
  Demonstrates strong task-agnostic fidelity for clearly defined socially desirable high-level persona
• GPT-4o ACCatom = 0.14 for Neutral-O persona in Questionnaire taskfinding0.730
  Very low atomic accuracy for neutral openness persona, illustrating difficulty of ambiguous neutral personas
• Secure control fine-tuning leaves moral susceptibility S near base levels for GPT-4o (-9%), GPT-4.1 (-20%), and Qwen3-235B (+2%)finding0.729
  Shows that susceptibility spike is specific to misalignment-inducing training signal, not generic fine-tuning
• GPT-4o fine-tuned on 6,000 insecure code completions became broadly misaligned, giving anti-human, violent, and deceptive answers on 50% of evaluation questions vs 0% for secure code controlsfinding0.725
  Key empirical result from Betley et al. 2025 that initiated persona vector research