Training-time intervention yields 63.4% blind-pair human preference at matched val-loss (1.2B params, 320 judgments, p = 1.98 × 10⁻⁵) [R]