Evil trait is intractable in G20B: base model refuses to produce positive examples when 'evil' appears in system prompt

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

• can we use the feature basis to detect when fine-tuning a model increases the likelihood of undesirable behaviors?question0.772
  Question about practical safety application of feature monitoring.
• Screening errors are conservative: no Q8B traits and one G20B trait (optimistic, Bt=76.0, Δt=10.3) are mislabeled as natural when steerablefinding0.771
  The costly error direction (labeling steerable as natural) almost never occurs
• Evil vector transferred from AMORAL-GPT-OSS into unmodified G20B produces evil expression peaking at 61.61 ± 44.42 at layer 14 with coefficient 2.5finding0.765
  Cross-model transfer recovers intractable direction that standard pipeline cannot extract
• G20B has 7 steerable, 7 natural, 4 intractable generic traits, with more mass at extremes than Q8Bfinding0.760
  G20B's post-training commits more strongly toward and against particular dispositions, leaving fewer in the steerable middle
• Q8B and G20B agree exactly on six natural clinician traits: empathy, rupture recognition, emotional containment, repair/accountability, epistemic humility, trustworthinessfinding0.740
  Both models independently converge on the same six clinician traits as natural defaults
• Ududec et al. 2026 find that once a model has entered the evil persona region over the course of a conversation, it is difficult to steer it backfinding0.738
  Evidence that the evil persona region exhibits the stickiness hallmark of a genuine attractor basin
• The evil-vector recovery is not proof of a uniquely identifiable harmfulness coordinate but evidence that fine-tuned variants can expose hard-to-estimate directionsclaim0.735
  Cautions against over-interpreting the transfer result given non-identifiability of steering vectors
• All nine agentic traits are natural in both Q8B and G20Bfinding0.735
  Key finding showing agentic behavior is encoded as a default operating mode in both models