Generation length shows weak correlation with ACCatom and RCatom (r = [0.20, -0.12])

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

• ICatom has very low correlation with generation length (r = [-0.31, -0.12])finding0.874
  Generation length does not substantially degrade internal consistency scores
• # Neutral shows weak correlation with diversity (ρ=0.05 decTest, ρ=-0.08 conTest)finding0.794
  Ablation result showing neutrals are not strong indicators of diversity
• Gen B (GPT-4o, neither extrovert nor introvert) exhibits ACCatom=0, ICatom=0.06 while ACC assigns it as alignedfinding0.753
  Motivating case study showing atomic metrics detect OOC behavior invisible to response-level metrics
• GPT-4o achieves near-perfect ACCatom=1.00, ICatom=0.97, RCatom=0.98 for High-N (emotionally stable) persona in Questionnairefinding0.740
  Demonstrates strong task-agnostic fidelity for clearly defined socially desirable high-level persona
• Most correlated neuron A/neurons/470 has correlation of only 0.18 with base64 feature A/1/2357 and responds to code, HTML labels, URLsfinding0.737
  Shows base64 feature is polysemantic at neuron level but monosemantic as learned feature
• Correlation r=0.999 between detection-adjusted logit difference and control logit increase across all 40 layer-strength configurationsfinding0.734
  Key quantitative evidence that detection signal is identical to global logit shift confound
• Pearson r = 0.91 between ACC and ACCatom across all modelsfinding0.729
  High correlation shows ACCatom measures similar underlying construct to ACC but with finer granularity
• VS-CoT with GPT-4.1 as generator achieves 45.9% accuracy on Qwen3-1.7B-Base fine-tuned on math benchmarks, the highest single resultfinding0.723
  Best performing VS variant for math synthetic data generation with GPT-4.1