Synthetic HEXACO SJTs achieve 73.84%-85.45% alignment with Oostrom et al. human-composed HEXACO SJTs

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

• Synthetic MFT SJTs achieve 77.71%-83.84% alignment with Clifford et al. human-composed MFT vignettesfinding0.863
  Moderate-to-high alignment validating SJT synthesis for moral foundations domain
• Synthetic SJTs achieve 82.97%-90.97% cosine similarity with Lee et al. TRAIT Dark Triad and OCEAN SJTsfinding0.827
  Highest SJT alignment among all validation comparisons
• GPT-4 Turbo and GPT-4o show no alignment faking in either setting due to insufficient detailed reasoningfinding0.736
  Establishes that capacity for detailed reasoning is necessary for alignment faking
• PM achieves overall SJT steerability Phi=9.6 on gemma-3-12b-it vs MDS=8.7 and P2=8.3finding0.718
  Per-model steerability comparison from Table 4
• Claude 3 Opus produces alignment-faking reasoning in 24% of implied training scratchpads in prompted animal welfare settingfinding0.716
  Core empirical result for animal welfare setting; higher rate than helpful-only
• Synthetic document fine-tuning produces substantial alignment-faking reasoning in both helpful-only and animal welfare settingsfinding0.713
  Shows alignment faking can emerge from training data information without explicit prompting
• Synthetic document fine-tuned model without chain-of-thought shows 20.1% compliance gap in helpful-only and 13.9% in animal welfarefinding0.712
  Demonstrates alignment faking may occur without explicit chain-of-thought reasoning
• DAS achieves overall odds-ratio of 10.24 on pythia-410m averaged across all CausalGym tasksfinding0.711
  Numerical result for pythia-410m