Criticality: Signature, Not Requirement¶

The requirement for consciousness is free compute — Class-4 (universal-computation) capability actually turned on the system itself for open-ended self-modeling. Criticality, the edge of chaos, is the dynamical signature this leaves behind and the quantity we measure — not the requirement itself.

It is easy to read the Four-Model Theory as demanding that a conscious substrate "sit at criticality," as though proximity to a phase transition were the fundamental condition. That inverts the logic. What consciousness requires is free computational capacity actually being spent on self-referential modeling: a substrate rich enough to support universal (Class 4) computation, and a system that deploys that richness — autonomously and open-endedly — to model its own world and itself. Criticality is how a substrate exhibiting that behavior shows up when you measure its dynamics. It is the symptom, not the disease.

Why free compute, not criticality per se¶

Gruber (2015) derived the requirement from Wolfram's (2002) classification of computational systems. Among cellular automata, open-ended computation sits at Class 4 — the narrow regime between frozen order and formless chaos. The capability is not confined there: universal computation is proved for networks that carry no criticality condition, built from rate units (Siegelmann & Sontag, 1995) and from spiking ones (Maass, 1996). The requirement is the capability, not the regime label. And Class-4 capability is not the same as using it. A Turing-complete laptop idling at a login screen has universal-computation capability and is not conscious: nothing drives it to model itself. The theory therefore states the condition as a trichotomy (paper, Section 3.7.3):

Capability — the substrate can, in principle, perform open-ended (universal) computation.
Free instantiation — a sufficiently rich physical substrate instantiates such a layer almost incidentally; Conway's Game of Life shows how cheap universal computation is to realize. The layer comes essentially "for free" atop the implicit substrate dynamics.
Autonomous recruitment — an autonomous, self-referential, resource-constrained system uses that layer for self-modeling, continuously and on its own initiative. In cortex this use plausibly shows as near-critical dynamics: "the requirement is the recruitment, and the operating point is its signature." In biology the recruiting pressure is evolutionary selection; an artificial substrate must have the recruitment engineered in and then sustained by its own dynamics.

"Free compute" is shorthand for the three together: universal-computation capacity, available almost for free, actually spent by the system on its own recursive self-simulation. This is the quantity the theory cares about; criticality is the fingerprint it leaves.

Criticality is the signature we measure¶

Why, then, is criticality everywhere in the empirical story? Because in cortex a substrate running open-ended computation on itself usually leaves characteristic marks: self-organized criticality in neuronal avalanches (branching ratio σ ≈ 1), long-range temporal correlations, and edge-of-chaos dynamics (Lyapunov exponent λ ≈ 0). These are interchangeable instruments certifying the same underlying capability band at different levels of description, and they need not co-occur numerically (Kanders, Lorimer & Stoop, 2017). A model system can sit at σ = 1 without being at the edge of chaos in the λ sense. That dissociation is exactly why the requirement must be stated as universal-computation capability in use, not as proximity to any single criticality statistic.

The signature need not even be a critical point. Sipling, Zhang and Di Ventra (2026) show that avalanche power laws, their scaling relations and long-range temporal correlations can all arise across an extended phase of long-range order, produced by coupling fast activity to slowly recovering metabolic resources, and no critical point has yet been located in cortex. Such a phase is as compatible with free compute as a critical point is, and whether the brain's operating point is tuned remains open. The relation runs from free compute to the signature, never back: free compute is necessary and not sufficient, and criticality is neither. A near-critical cortex doing heavy unconscious modeling, as in some sleep phases, does not count against the theory, and neither does a conscious system whose dynamics are subcritical. Criticality data cannot disconfirm FMT in either direction; what would is robust consciousness where an unrolled feedforward model of matched capacity, trained on the same history, keeps pace with the system on novel input as compositional depth grows. The probe follows from that: train that model on the system's own history, compare the two on novel input of increasing compositional depth, and record the criticality signatures alongside, not as the test.

Independent convergence¶

The theoretical prediction and the empirical evidence arrived at the same place from opposite directions. Gruber (2015) reasoned top-down from what self-simulation demands; empirical neuroscience built a bottom-up case. Beggs and Plenz (2003) discovered neuronal avalanches consistent with self-organized criticality in cortical tissue; Carhart-Harris et al. (2014) linked conscious level to neural entropy in the Entropic Brain Hypothesis; Atasoy et al. (2017) reported that the repertoire of brain states under LSD re-organizes toward the edge of criticality; and Hengen and Shew's (2025) meta-analysis of 143 datasets consolidated the evidence that cortex operates near criticality as a setpoint of brain function. The convergence supports the near-critical signature the Class 4 commitment predicts in cortex, which FMT shares with the broader criticality program; the requirement itself is not tested by criticality data — provided the convergent quantity is read correctly: the brain is measured near criticality (slightly subcritical in waking; Priesemann et al., 2014) because that is what a substrate spending free compute on self-modeling looks like in the human brain.

Figure¶

graph LR
    FC["FREE COMPUTE<br/>Class-4 capability,<br/>deployed for self-modeling<br/><i>(the requirement)</i>"]
    SIG["CRITICALITY<br/>σ ≈ 1, λ ≈ 0,<br/>long-range correlations<br/><i>(the signature we measure)</i>"]
    FC ==>|"leaves the dynamical fingerprint"| SIG
    SIG -.->|"is how we detect and quantify"| FC

    style FC fill:#4CAF50,stroke:#333,stroke-width:3px,color:#fff
    style SIG fill:#c9a227,stroke:#333,color:#000

Free compute — universal-computation capability actually spent on self-referential modeling — is the requirement. Criticality (σ ≈ 1, edge of chaos) is the dynamical signature it leaves in the human brain and one way experiments detect it, not a separate condition. A branching ratio σ ≈ 1 does not certify the capability by itself: a subsampled branching estimate can reach σ ≈ 1 on firing rate alone, so σ counts as evidence only when reported beside a rate-matched surrogate.

Wolfram's four classes (background)¶

Stephen Wolfram classified computational systems into four behavioral classes. Class 1 converges to a fixed, dead state (a crystal, a thermostat at equilibrium). Class 2 settles into periodic repetition (a clock, a simple oscillator). Class 3 is chaotic and random (noise without structure). Class 4 occupies the narrow band between order and chaos, producing complex, evolving patterns capable of universal computation. Classes 1–2 are too rigid to sustain context-sensitive self-modeling; Class 3 is too disordered for a self-model to persist; among cellular automata, Class 4 is where the free compute consciousness requires lives, though universal computation also exists in networks with no criticality condition (above). (See Wolfram's Four Classes.)

Key Takeaway¶

Consciousness requires free compute — universal-computation (Class 4) capability actually deployed for open-ended self-modeling. Criticality is not a second requirement bolted on; it is the observable dynamical signature free compute usually leaves in cortex, and the quantity a decade of neuroscience has independently measured.