My earlier work asked how artificial lexicons change human learning and communication. The same experimental instinct now guides questions about model training: control the code, control the evidence, and measure what is learned under a defined resource constraint.
The question that remains
Recovering a fact from a model’s activations does not by itself explain how the model uses it. Behavioral tests, interventions and suitable controls matter. Current working notes revisit these distinctions; no general law of human–machine information capacity is claimed.
