PersonaQA closed set, exact %, s(v)=log p(v|a_i) - lam*log mean_j p(v|a_j) over all 100 personas; paired persona-bootstrap CI for lam=.75 vs 0
  original  0:11.8 0.25:12.3 0.5:13.3 0.75:14.7 1.0:14.3 1.25:4.0   d(.75-0) +2.8 [+0.7, +5.0]
  improved  0:22.5 0.25:23.5 0.5:25.7 0.75:26.5 1.0:23.2 1.25:15.8   d(.75-0) +4.0 [+1.5, +6.5]
  grpo      0:16.2 0.25:18.8 0.5:21.3 0.75:23.7 1.0:20.8 1.25:13.5   d(.75-0) +7.5 [+4.7, +10.3]
Synthetic two-person contexts (2-entity items): binding error (answer = other person value), raw vs ratio against the other person activation
  original  raw 14.5%  ratio 7.0%  d -7.5 [-13.1, -2.0]
  improved  raw 12.5%  ratio 7.5%  d -5.0 [-10.7, +0.5]
  grpo      raw 15.0%  ratio 11.5%  d -3.5 [-9.1, +2.1]
