Can a small network learn new digits without forgetting old ones?

A user describes a Claude-assisted Split MNIST experiment: a small neural network learns handwritten digits in order—0/1, 2/3, 4/5, 6/7, then 8/9. Each stage uses only that pair’s real training images, with testing on all ten digits after the final stage. The user says the setup starts with two training passes per task and reports results averaged across three random seeds.