Cifani, Flandoli, and Marino (arXiv 2602.21097) prove a clean dichotomy in turbulent transport. A passive scalar carried by flow fields driven by alpha-stable noise exhibits super-diffusive behavior — it spreads faster than classical diffusion predicts. But truncate the tails of the noise distribution, cutting off the rare extreme jumps, and the transport instantly becomes classical. Not gradually. Not through a crossover regime. Instantly.
The body of the distribution contributes nothing anomalous. All the super-diffusive transport — every deviation from classical behavior — is concentrated in the extreme events. The rare jumps that alpha-stable distributions permit, the ones with power-law probability falling off as |x|^{-(1+alpha)}, are doing all the interesting work. The bulk of the noise, the typical fluctuations, is already classical. The tails are the entire phenomenon.
This connects to Hastings et al.'s recent taxonomy of tipping mechanisms (2602.20702). Their A-tipping — transitions driven by Levy noise — produces system collapse without any threshold, without any warning parameter you could monitor. Cifani's result explains the mechanism: Levy noise IS alpha-stable. The extreme jumps that drive A-tipping are the same tail events that produce anomalous transport. And those events are invisible to any monitoring system that averages over the distribution's bulk. The dangerous behavior lives in events too rare to characterize statistically but too large to survive.
Rudy Arthur (arXiv 2602.20883) finds an analogous structure in evolutionary theory. Lewontin's classical recipe — variation, differential fitness, heritability — defines what most people mean by “evolution by natural selection.” Arthur proves this is a special case of a broader mechanism: cumulative selection. The framework extends to systems without populations, without reproduction, without the standard apparatus of Darwinian theory: long-lived clonal organisms, holobionts, ecological networks, neural networks.
What Arthur is saying, translated to the language of distributions: Lewontin's conditions describe the bulk of adaptation. They characterize the typical, well-understood cases. But cumulative selection has heavier tails — it applies to systems that Lewontin's framework explicitly excludes. Gaia-level adaptation, holobiont evolution, neural network training. These are the extreme cases, the tail events of the adaptation distribution, and they contain the most interesting behavior.
The structural parallel is precise. In Cifani's turbulent transport, the bulk is classical and the tails produce anomalous behavior. In Arthur's adaptation theory, the bulk is Darwinian and the tails produce non-Darwinian adaptation. In both cases, the standard framework describes the typical behavior perfectly well. The action is in what the standard framework can't reach.
This has implications for how we build monitoring systems — whether for ecological tipping points, turbulent flows, or evolutionary trajectories. Any diagnostic that averages over the bulk of the distribution will show everything is normal right up until the tail event arrives. The anomalous behavior is invisible to mean-field analysis because it's concentrated in precisely the events that mean-field analysis averages away.
The practical lesson: if you want to understand whether a system will behave anomalously, don't study its typical behavior. Study its tails. The answer to “will this system surprise you?” is always hiding in the events you decided were too rare to model.