Metric score (1–100) next to the human win-rate — how often each fold was chosen
as more complex when shown. Numbers refresh weekly from the collected votes.
Sort by
Structure
Fold
aa
Score
Games
Human win-rate
Community & calibration
How people have engaged since launch, and how the metric responds to their votes.
User engagement per hour
Votes and score-ratings each hour since launch, with the cumulative distinct-user count.
votes / hrratings / hrcumulative users
Does the metric improve as people vote?
Weights fitted on the first N votes, then scored on the held-out remainder — the share of
later votes where the fitted metric picks the same "more complex" structure the human did. Dashed line is
today's weights on the same held-out votes.
fitted on first N votescurrent weights
Reconstructed by training on the first N votes — not a record of what was live
(past weights weren't stored; logging is now on, so real history accrues from here).
How each component's weight would evolve
Fitted weight of each of the ten components as the training set grows to all votes.
Shared scale (0 → 0.30); the dot marks the value at the full vote set.