capacity-matching

Capacity-matching is the rule that a model or proxy predicts a human learner only as far as it shares the human's limits — the boundary where the two part ways is the boundary where their capacities diverge.

An AI student's learnability has never been correlated with a human's, because the two part ways at the points where their capacities differ: the AI may have a larger working memory (so a high-element-interactivity concept that blocks a human is trivial for it), a different perceptual system (so a concept that leans on visual pattern is easy for one and hard for the other), or a vastly larger sample appetite (so a concept that needs a thousand examples is a weekend for the AI and a semester for the human). The proxy predicts the human only where the proxy's limits match the human's limits — and the more the proxy exceeds the human, the smaller the matching window becomes. A domain-matched student model (one that shares the human's capacity constraints) should predict human learning better than a generic one, because the capacity-matching window is wider.