A local guarantee that the learned score function (gradient of log-density) accurately matches the true score at convergence.