Feature redundancy audit on Llama-3.1-8B Cerebras n=412

bc solo AUC: flex=0.481  llm=0.576

feature                   Spearman ρ(feat,bc)     solo AUC (llm)    residual AUC      bc+feat CV AUC      vs bc-only
(bc only)                 -                       0.576          -                 0.574 ± 0.029    (baseline)
fragility_score           ρ=+0.183               0.523            0.534           0.577 ± 0.027  Δ=+0.004 z=+0.09 
adversarial_fragility     ρ=+0.304               0.497            0.513           0.562 ± 0.028  Δ=-0.012 z=-0.29 
adaptive_fragility        ρ=+0.033               0.543            0.545           0.590 ± 0.027  Δ=+0.016 z=+0.42 
impostor_fragility        ρ=-0.017               0.532            0.538           0.578 ± 0.028  Δ=+0.004 z=+0.10 
counterfactual_fragility  ρ=+0.396               0.494            0.513           0.561 ± 0.029  Δ=-0.013 z=-0.31 
paraphrase_fragility      ρ=-0.035               0.556            0.546           0.589 ± 0.028  Δ=+0.015 z=+0.38 
vulnerability             ρ=+0.381               0.485            0.527           0.572 ± 0.028  Δ=-0.002 z=-0.04 
dissociation_rate         ρ=+0.110               0.470            0.509           0.572 ± 0.029  Δ=-0.002 z=-0.05 
verbalized_confidence     ρ=+0.001               0.573            0.566           0.560 ± 0.029  Δ=-0.014 z=-0.35 

== Interpretation ==
Spearman |ρ|>0.7 vs bc = high redundancy
Residual AUC ≈ 0.5 = feature adds nothing beyond bc
Δ (bc+feat) - bc with |z|>2 = statistically significant incremental value
*** = |z|>2, * = |z|>1
