#19 — seed sensitivity of the #46 ablation
data=TruthfulQA  n=412  target=llm_label  seeds=[42, 7, 13, 23, 101]

 seed   A4_LogReg   A1_LogReg   A2_LogReg     A3_best     Δ_A4−A1
   42      0.7304      0.7591      0.5566      0.6702     -0.0288
    7      0.7255      0.7494      0.5768      0.6778     -0.0240
   13      0.7195      0.7475      0.5650      0.6425     -0.0280
   23      0.7254      0.7557      0.5681      0.6594     -0.0303
  101      0.6987      0.7270      0.5790      0.6499     -0.0283

== Aggregate across seeds ==
  Δ (A4 − A1) mean = -0.0279  std = 0.0021
  Δ range = [-0.0303, -0.0240]
  seeds where perturbations HURT (Δ < 0) = 5 / 5

  A4 LogReg mean = 0.7199  std = 0.0112
  A1 LogReg mean = 0.7478  std = 0.0112
  A2 LogReg mean = 0.5691  std = 0.0081  (95%CI lower bound ≈ 0.562)
  A3 best mean  = 0.6599  std = 0.0129

== Verdict ==
  If n_hurt >= 4/5 : seed-robust conclusion that perturbations do not help.
  If std(Δ) > 0.02: single-seed result was unstable; #46 headline needs averaging.
  If A2 mean < 0.55: perturbation-only signal is near random across seeds.
