benchgap
Calibration

HealthBench (raw) → ARC-AGI-1

ARC-AGI-1 is estimated from HealthBench (raw) with a linear curve fitted on 6 models measured on both: y = 0.2810·x + 0.7917, R² = 0.31, cross-validated error 5.0 pp. It is used for 0 estimates.