Each panel: probability Chibany asks the same friend again, split by yesterday's bento (tonkatsu vs hamburger) and whether the friend brought their usual box (a common transition) or the other color (a swap, the rare one).
Tonkatsu and hamburger lines that stay parallel = a pure reward effect (model-free);
lines that cross = a tonkatsu×swap interaction (model-based). Slide w to mix the two — the middle panel is the human-like blend.
Knobs: w = plan (model-based) vs. cache (model-free) · β = how sharply each agent picks the higher-valued option (softmax inverse temperature) · α = its learning rate (same α as the chapter's Q-learning update).