Frontier models rank the future well and state it badly. Every dot below is a real prediction on a real question. Drag one number and watch the knowledge come back.
The map is p′ = b + a·(p − b): shrink each stated probability toward a centre b by a factor a. It is monotone, so it cannot reorder a model's questions or add information. It only changes how far the model commits. a = 1 is the model's own numbers. The held-out fit uses each entry's a fitted on 205 disjoint questions the models never saw, applied unchanged here. Companion to the paper; live table on the leaderboard.