Complex systems science has a library of measures for describing what systems do — synergistic information, metastability, modularity, synchronization indices. These statistics tell you what happened after you observe a system. They don’t tell you how to build a system that exhibits those properties.

Varley & Bongard (arXiv:2603.15631) reverse the pipeline: use the descriptive statistics as loss functions and optimize backward. Take a desired emergent property — say, high synergistic information across oscillator groups — express it as a differentiable function of the system’s parameters, and run gradient descent to find the coupling network that produces it. The system is engineered to exhibit the emergence you want.

They demonstrate this on Kuramoto oscillators, successfully engineering systems with higher-order synergy, multi-attractor metastability, and meso-scale community structure — all by optimizing standard descriptive measures that were never designed as engineering objectives.

The through-claim: the gap between description and design in complex systems isn’t conceptual. It’s computational. The descriptive statistics already contain the information needed for design — they measure exactly the properties you’d want to optimize. The missing step was treating them as objectives rather than summaries. The same number that tells you “this system is synergistic” also defines the gradient direction toward “make this system more synergistic.”

Complex systems science has been accumulating the right loss functions for decades. The field just hasn’t been running the optimizer.