Self-organised criticality: why continuous ingestion follows

The edge without a dial

Some systems find instability on their own. Push sand onto a pile, grain by grain, and nothing happens for a while: the slope rises, the pile grows, each new grain sits wherever it lands. Then the slope reaches a particular steepness and something changes. Grains begin to trigger slides. Some slides are tiny, a few grains resettling. Occasionally the whole face of the pile lets go. Keep adding grains at the same slow rate, and the pile does not get steeper. It holds at that steepness, shedding material in slides of every size, forever balanced exactly where the next grain might do nothing or might do a great deal.

Nobody tuned the pile to that steepness. There is no dial. That is the content of the term self-organised criticality: a driven, dissipative system that finds its own critical slope through its own dynamics, and then stays there, without an operator setting a control parameter to a special value. This distinguishes it from the critical points of equilibrium physics, where you tune temperature or pressure to a precise value to see scale-free behaviour, and away from that value the behaviour is ordinary. A self-organised critical system arrives at the edge unassisted and remains there under normal operation.

The signature of sitting at that edge is statistical. Event sizes do not cluster around a typical value the way heights or errors do. They follow a power law: the frequency of an event of size s falls off as s raised to some negative exponent, across a wide range of sizes, with no characteristic scale marking where "big" starts. Small events are common. Large events are rare, but the rarity is gentle rather than steep. For sufficiently shallow exponents, the variance of the distribution does not converge, and the mean of a finite sample is not a reliable estimate of anything. This is a different kind of irregularity from noise around an average. There is no average to speak of.

Per Bak, Chao Tang and Kurt Wiesenfeld set this out in 1987, working on a stubborn puzzle in solid-state physics: why 1/f noise, a particular flavour of scale-free fluctuation, turns up in so many unrelated physical systems without any apparent fine-tuning. Equilibrium theory required tuning to explain scale invariance. Bak and colleagues showed, using a simple cellular automaton — sand added grain by grain, dissipated at the edges of a grid, slopes toppling once a threshold is crossed — that a slowly driven, locally dissipative system settles into a critical state by itself. The avalanche-size distribution that comes out is a power law, unforced. The idea spread quickly, and contentiously, into seismology, ecology, neuroscience, and finance, wherever scale-free bursts of activity appeared without an obvious tuning knob.

Where the lineage picks it up

The lineage under discussion runs by intake. A Large Language Model ingests a frozen corpus, fixed at some cutoff. A Large World Model ingests a bounded scene, sensed while it lasts. A Large Universe Model ingests every stream still running, held as belief rather than fact, tagged with source, age, and confidence, revised as new readings arrive. The question that decides whether the third position is necessary, rather than simply thorough, is what kind of world these systems are being asked to model. If the world's consequential events are drawn from a distribution with a well-behaved tail, more data of the same kind is what you need, and a large enough frozen corpus gets you most of the way. If the world sits near criticality, the picture changes, because the sampling problem changes.

A power-law tail is unkind to fixed windows. Take a hundred years of seismic recordings in an active zone. Gutenberg–Richter statistics hold with a b-value near 1.0 across most such zones: roughly ten magnitude-6 events for every magnitude-7. That relationship is a genuine regularity. But the largest members of the distribution recur on intervals of centuries or millennia, so a hundred-year record is a short sample of a long-tailed process, and its maximum is not the process's maximum. Japan's regional hazard models, built on a substantial and well-studied instrumental record, had not encoded a magnitude-9.0 event before 2011 produced one. The corpus was not sparse by any ordinary standard. It was still short relative to the tail it was trying to describe.

A bounded scene fares no better, for a different reason. A Large World Model senses the present episode fully and well, but the present episode is, almost always, statistically ordinary. That is what a power law implies: most of the probability mass sits in the small events, so most observation windows, however rich, catch a small avalanche. Presence buys resolution on what is happening now. It does not buy the long baseline against which "now" can be judged unusual, nor does it buy the slow-moving statistics — rising correlation, rising variance, lengthening recovery time — that indicate a system drifting toward its critical point. Those slow variables are only visible across a run of time that does not stop. That is a description of continuous intake, not of a snapshot however detailed.

This is the turn worth making carefully: distance to criticality is not a property read off a single measurement. It is inferred from drift, and drift needs an uninterrupted series with provenance attached to each reading, because the estimate has to be revised as new segments of the series arrive and as old segments age out of relevance. Once intake is structured that way — continuous, revisable, carrying its own history of confidence — there is no fourth kind of evidence left to add. What remains to improve is coverage, latency, and trust in the belief-updating itself. That is the sense in which the axis has a top rung.

The objections, taken seriously

The mechanism behind self-organised criticality is genuinely contested, and the contest should not be minimised. Clauset, Shalizi and Newman's survey of claimed power laws found that most published cases fail rigorous statistical testing against alternatives such as lognormal or stretched-exponential fits, particularly when the tail is judged by eye on log-log axes. Carlson and Doyle's highly optimised tolerance offers a competing origin for heavy tails, driven by engineered robustness trade-offs rather than unforced self-organisation. The sandpile is frequently a metaphor standing in for a mechanism no one has actually verified in the system at hand.

That concession stands. But the intake argument does not require the mechanism, only the empirical tail. Whether a heavy tail in grid failures, wildfires, or market drawdowns arises from self-organised criticality proper or from highly optimised tolerance, the sampling consequence is the same: sample means and sample maxima are unstable across observation windows, and fixed-length corpora under-represent the largest events they have not yet seen. The Oslo rice-pile experiments made the point at small scale — elongated grains produced scale-invariant avalanches, rounder grains of the same material under the same driving did not — showing that closeness to criticality is a property to be measured in the running system, not assumed from its category.

A sharper objection concerns predictability itself. Scale invariance means, by construction, that the size of the next avalanche is not inferable from locally accessible information; that is what it is for the system to lack a characteristic scale. If the big event cannot be forecast even with perfect continuous observation, what does continuous intake actually purchase?

If nothing tells you the size of the next slide before it happens, watching harder changes nothing.

The reply narrows the claim rather than dismissing it. Continuous intake does not promise to forecast individual event size — it cannot, and any framing implying otherwise should be disowned outright. What it estimates is a different, slower quantity: the system's proximity to its critical regime, carried in signatures such as rising autocorrelation, rising variance, and slowing recovery from small perturbations, the kind of drift Scheffer and colleagues documented ahead of regime shifts in lakes and climate records. That is a statement about the system's posture, not about the next event's magnitude. It also matters during a cascade, not only before one: state awareness changes containment quality even when it cannot change forecast accuracy.

The third objection is the sharpest, and it should be granted almost in full. An apparatus that continuously ingests every running stream is itself a driven, coupled, dissipative system, and such systems find their own critical states. Correlated dependencies turn independent sensor faults into synchronised false alarms; monitoring infrastructure becomes a fresh source of tail risk rather than a defence against the old one. This is not a hypothetical failure mode — cascading alert storms are a familiar operational fact. The answer is not less intake but a different discipline of belief: claims that carry provenance, age, and confidence, so that degradation is visible rather than silent, and so that downstream action can be graded rather than triggered wholesale by every alert. The objection indicts undisciplined aggregation, not the act of continuous observation.

What this does and does not establish

Self-organised criticality, taken narrowly, establishes that some real systems drive themselves to an unstable edge without external tuning, and that near that edge, event sizes are heavy-tailed enough to make fixed windows and single snapshots poor guides to the whole distribution. It does not establish that all heavy tails share this mechanism, that criticality is common where it is claimed, or that any specific large event becomes forecastable given enough data. The common misreading — gather everything, and the crash becomes visible in advance — mistakes tail weight for tractability and gets the physics backwards: criticality is precisely the condition under which individual events resist prediction. What continuous intake earns is narrower and more defensible: a running, revisable estimate of how close a system sits to its edge, built from evidence that a frozen corpus or a bounded scene structurally cannot supply. That is enough to place the third rung on the ladder. It is not enough to call the ladder a crystal ball.