Each point = one of the 26 resolved events. X = production p_yes for outcomes[0]. Y = single-binary Brier loss. Green = outcomes[0] won; red = outcomes[0] lost. Size = n_outcomes (bigger circles = harder multi-outcome events). Hover for question + winner + rationale.
Read the picture: green dots on the right (high p_yes for the actual winner) -> low Brier (good); red dots on the right (high p_yes when outcomes[0] LOST) -> high Brier (bad confident-and-wrong). Green dots on the left (low p_yes when outcomes[0] WON) -> also high Brier. The shape of the curve shows where our model's confidence helped or hurt.