I'd be fascinated to know how flexible these are as neural ODE ansatze.

Looks to me they could implement arbitrary logic, but could we prove it.