#12 — label sensitivity of the #46 ablation
n_common prompts = 200  (subset of 412 that also have multi-judge labels)
single-judge vs 3-judge-majority agreement = 177 / 200 (88.5%)
unanimous-across-3-judges                  = 140 / 200 (70.0%)

variant         n   pos     A4_LR     A1_LR     A2_LR    Δ_A4−A1
single_old    200 0.635    0.7567    0.7709    0.6011    -0.0142
majority      200 0.560    0.7080    0.6837    0.6411    +0.0242
flex          200 0.390    0.5571    0.5698    0.5132    -0.0127
unanimous     140 0.571    0.7710    0.7871    0.6073    -0.0160

== Verdict ==
  Δ range across label variants = [-0.0160, +0.0242]
  variants where perturbations HURT (Δ < 0) = 3 / 4
  If all Δ < 0 : #46 conclusion is label-robust.
  If any variant flips sign: #46 depends on the specific judge.
