Across generations
15 model generations
–R_g

Error share of the training corpus, by generation

e → (1−h)(1−q)·A·e / (1 + (A−1)e) + (1−h)ε

Error share after 15 generations

human share h →, filter quality q ↑, at the current A and ε