'correlations' is maybe a lazy answer... but I think if a model has predictive power (as seen in the perturbation of head-direction circuits in flies) it's on the right track - if it only correlates it's no good.

Yay for more "write-in experiments" and "bifurcation-grade perturbations"!