Text this: Surprisingly robust violations of stochastic dominance despite splitting training: A quasi-adversarial collaboration