Leaderboard

Time-weighted, not a flat average.
Time-weighted: D = Σ(d_i × t_i) / Σ t_i. Each puzzle's contribution is scaled by the time you spent on it.
A 30-second hard puzzle (d=8) and a 2-second easy one (d=1) flat-average to 4.5, which makes the run look way easier than it felt. Time-weighted: (8×30 + 1×2)/32 = 7.6. This method lets the 30 sec difficulty feel more significant.

Cluster median, not your own time.
Difficulty per puzzle: d_i = 10 × (m_c − 1500) / 5500, where m_c is the global median response time for that puzzle's cluster (e.g. mul_hard_large). Your speed never enters the formula, only the puzzle's reputation across all players. So 12×8 has the same d whether you did it in 2s or 9s.

Median, not mean (for m_c).
One AFK player sitting on 12×8 for 60 seconds would drag the mean up and make the puzzle look harder than it is. Median ignores them. We compute m_c as the median of all correct, non-practice attempts in the cluster.

Time capped: t_i = min(response_ms, 15000).
Same outlier problem on the run side. If you alt-tabbed for 30s, that puzzle wasn't 30-seconds-hard, you just left. Capping stops one distracted moment from dominating Σ t_i.

Floor 1.5s, ceiling 7s in the d formula.
1.5s is roughly the fastest a human can read-and-type, so anything that fast scores 0. Around 7s is where puzzles start actually hurting, so that's the 10. Linear in between. Picked from eyeballing real attempt data, clamps to [0, 10] so weird clusters can't blow past the scale.