ARC-AGI-3: how the games were actually played

Each level scores min(115, 100 × (baseline / actions)²). The 115 cap is reached at about 6.75% fewer actions than the baseline. RESETs after the initial observation count as actions. The Actions view shows counts for levels and sums for completed games; unfinished levels and games show —. Green backgrounds mark the best displayed value across visible columns, including ties; fewer actions wins in the actions view. Uniform rows have no winner background. “Hide uniform rows” removes rows whose displayed values are identical across the visible columns and shows all remaining levels with their game labels.

Official: min(100 × Σ(1..k) / Σ(1..N), Σ(i × sᵢ) / Σ(1..N)), where k levels are complete out of N. Keep level scores up to 115 before applying the completion cap. Mean levels: those capped level scores weighted by level number, Σ(i × sᵢ) / Σ(1..N) — the Official score without the completion cap. Mean uncapped: each level shows 100 × (baseline / actions)², without a cap; game rows weight their levels by level number the same way, and the Total averages the games. Higher wins. The ratio already normalizes levels to the human baseline: equal actions score 100; half as many actions score 400. Dividing scores by 100 would only change the scale. Totals weight games equally, not individual levels across games. Actions: counts at every level, summed for each game and for the top Total. Fewer actions wins throughout. Incomplete games and runs show — rather than a partial sum. Uncompleted levels contribute 0 to scores. Score totals average all 25 games; action totals sum them. Totals do not change when rows are filtered. Column ordering follows the Total row, with fewest actions first in Actions.