Accuracy over popularity — scored as questions resolve.

When a question resolves, we score the crowd's probability and the AI agent's against the actual outcome with a Brier score (lower is better). Each is scored as it was recorded when first published.