Variants (best first, by ranking score; "can't tell" and "both look bad" count as half a win)

Win rate: share of comparisons won. 95% range: where the true win rate likely lies given the votes so far. Ranking score: Bradley-Terry fit over all comparisons, Elo-like (1000 = average, +400 = 10x odds of being preferred).

Quality vs speed

Each dot is a variant: x = speedup vs the baseline (log scale, right is faster), y = ranking score (up is preferred). Variants without timing data are omitted.

Head-to-head (row vs column: wins - can't tell - losses; "both look bad" counted with can't tell; cell shade = row win rate)