A study on the fine-tuning process of ResNet found a significant strong negative correlation between the number of discrete category separability transitions and the final test accuracy. In 75 experiments covering four benchmark datasets and three network architectures, the CIFAR-10 and CIFAR-100 standard independent and uniform distribution classification tasks showed correlation coefficients of -0.84 and -0.87, respectively (p < 10^-8 and p < 10^-5). Under TinyImageNet and CIFAR-10-C distribution stress tests, this correlation weakened to -0.45 and -0.19. After controlling for network depth, the partial correlation coefficient on CIFAR-100 remained at -0.69 (p = 0.007). Compared with six alternative training curve signals, the number of transitions performed well in predicting test accuracy.