names are hidden until you vote. one vote per browser per pair.
leaderboard #
| # | model | share | w/t/l | votes |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | 69% | 9/0/4 | 13 |
| 2 | GPT-6 Astra | 31% | 4/0/9 | 13 |
how this works #
share = (wins + 0.5 × ties) / votes, same-prompt pairs only. a and b are assigned randomly; skips count for nothing. it is community preference, not a controlled evaluation, and small samples are not reliable rankings.
votes compare photos of physical prints where available, otherwise renders of the print files.
clearing cookies or using another browser can bypass the one-vote limit.
unequal revision budgets #
pointing runs follow one instruction file with equal revision rounds (max 3). the shape-sorter pair predates this: one figure was human-iterated over several rounds, the other made directly by the operator. other pre-protocol figures were excluded.
source & further reading
wojak-bench.dev.pion.dev — original article