The statistics we use to describe emergent properties — synergistic information, metastability, meso-scale structure — can be repurposed as loss functions. Feed them to a gradient descent optimizer, and it discovers the micro-scale interactions that produce those macro-scale properties. Description becomes design by reversing the direction of the same measurement.
This inversion is not a trick. It's a claim about the nature of descriptive statistics: if a statistic captures the essential structure of an emergent property, then optimizing for that statistic should produce systems exhibiting that property. The fact that this works — demonstrated on coupled oscillator systems engineering specific forms of complexity — means the statistics were doing more than summarizing. They were encoding the generative constraints.
Note (2026-11-14; read against the paper's full text): the essay presents the inversion as a general claim about descriptive statistics, that a statistic which captures the structure of an emergent property will, when optimized, produce systems with it. The paper's claim is more conditional and technical: given a differentiable descriptive statistic and a differentiable system, computed in a single computational graph, the gradient can be backpropagated to the parameters that define the system. It is demonstrated on Kuramoto systems of coupled oscillators as a test bed, for properties such as higher-order synergistic information, multi-attractor metastability, modularity and integrated information, with multi-objective cases handled by linear combinations of loss functions or by alternating epochs, and it says the pipeline can accommodate constraints on the system. The essay's claims that the statistics were 'encoding the generative constraints' and that the descriptive vocabulary is the design space are the essay's own readings.
The deeper structure is in what makes this different from specifying the emergent property directly. You can't hand-design a system to exhibit synergistic information. The micro-scale interactions that produce it are too numerous and interdependent to enumerate. But you can measure synergistic information in an existing system, express that measurement as a differentiable function, and let the optimizer search the space of micro-scale interactions for configurations that maximize it.
The description wasn't just a measurement. It was a compressed specification of the mechanism. The difference between description and design was not conceptual but directional — the same instrument points backward (what does this system have?) or forward (what should a system have?). The mathematics is identical. Only the causal arrow changes.
This has an uncomfortable implication for science. If descriptive statistics can serve as generative specifications, then the choice of what to measure is also the choice of what can be engineered. The statistics we don't develop — the emergent properties we lack vocabulary for — are precisely the properties we can't design. The descriptive vocabulary IS the design space.