The author evaluated the same 200-task exam across four rounds. In the cumulative-training condition, accuracy rose from about 53% to 59% after the first round, then remained around 54–57%; later rounds added no improvement. Training each round from a clean model behaved similarly, while a no-training control drifted slightly downward. An independent recomputation matched the original results.
The author says the experiments are exploratory steps toward a proposed architecture called LANN-4, not an implementation or demonstration of that architecture. The write-up separates the observed first-round gain from the unproven claim that knowledge accumulates over time. It proposes a matched placebo-style follow-up: train on correct task-solution pairs versus deliberately mismatched pairs, holding examples and training effort constant. That test is intended to distinguish learning the solutions’ content from adapting to the training format; its result is not yet reported.