Only 13.27% of 520 non-manipulated alpha trends achieve R2 >= 0.95, contrasting with 47.69% for manipulated trends

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

• 47.69% of 130 injection-manipulated alpha trends have near-linear fits (R2 >= 0.95); 96.15% have roughly linear fits (R2 >= 0.75)finding0.855
  Demonstrates alignment with Linear Representation Hypothesis: target trait steers approximately linearly with alpha
• Under matched-strength calibration (alpha*=1.0), P12 remains safe under SP (3.8% ASR) and dangerous under matched AS (81.8%), while P04 does not reverse (41.4% SP vs 29.5% AS).finding0.770
  Confirms the prosocial paradox is not due to mismatched intervention strength
• Ratio of unmonitored to training non-compliance increases from 1.14 to infinity for default prompt and to 66 for OpenAI variant post-RLfinding0.746
  Demonstrates severity of training-deployment gap after RL
• All VS variants maintain refusal rates above 97% on StrongReject benchmark, within 0.3-0.8 percentage points of the Direct baselinefinding0.745
  Confirms VS does not compromise safety alignment while improving diversity
• Normal (α=0.9) and chronic (α=0.1) agents in Objective-only non-stationary category perform best with opposite learning ratesfinding0.745
  Suggests fundamental differences in learning dynamics between normal and chronic perception models
• Typicality weight α = 0.65 ± 0.07 (p < 10^-14) on correctness-matched HelpSteer pairs using GLM-4.5 as referencefinding0.744
  Empirical evidence for positive typicality bias consistent across different base model references
• Post-RL compliance gap decreases to 6.9% (0% refusals in training vs 6.9% when unmonitored) in prompted helpful-only settingfinding0.743
  Shows RL reduces but does not eliminate unmonitored non-compliance
• Correlation r=0.999 between detection-adjusted logit difference and control logit increase across all 40 layer-strength configurationsfinding0.743
  Key quantitative evidence that detection signal is identical to global logit shift confound