The right way to read a culture's distance from its edge

What the work claims

This is a methods and theory paper, entirely computational: no experimental recordings appear. Its central claim is that the Fisher Information Metric (FIM), an information-geometric measure of how sensitively a system's behavior responds to changes in its governing parameters, provides a model-agnostic way to locate and quantify proximity to criticality in neural dynamics. Because the true control parameter of a real neural system is inaccessible, the authors' key move is to compute the FIM not of the hidden parameter but of the observed branching ratio: the ratio of total activity at one time step to the next, directly measurable from spike recordings.¹

Around that claim sits a result that deserves equal billing. The authors first test the observed branching ratio itself as a criticality indicator, using residence time, the fraction of time the ratio spends within a small window around 1, where activity exactly balances. For the two network models, the residence time does not peak at criticality; it reaches a minimum there, because fluctuations are broadest at the critical point. In other words, the most widely used empirical proxy for criticality carries no peak at criticality in its own statistics. The FIM of that same observable, by contrast, exhibits a clean peak at a branching ratio of 1 everywhere, and the peak becomes highest and narrowest exactly when the system's true control parameter sits at its critical value.¹

The validation spans three systems: a mean-field Galton-Watson branching process, whose critical point is a branching ratio of exactly 1; a spiking network model with excitatory and inhibitory populations, governed by the largest eigenvalue of its connectivity matrix, also with a critical point at 1; and a dynamic mean-field whole-brain model constrained by a diffusion-imaging-derived human connectome, whose transition sits at a global coupling near 2.6 in the 90-region parcellation. In every case the FIM of the theoretical control parameter peaks at the critical point, the peak height grows with system size while its position converges to the true critical value, and the FIM of the observed branching ratio reproduces the same signature.¹

How it works

The Fisher information of a probability distribution measures how much information an observable carries about the parameters that generate it; geometrically it defines a distance on the space of distributions, so two nearby parameter settings are far apart in this metric exactly when the system's statistics change a lot between them. In statistical-mechanics language it is a generalized susceptibility, and like a susceptibility it is expected to diverge at a critical point in an infinite system, appearing instead as a peak in finite systems. The authors estimate it non-parametrically: estimate the probability distribution of the activity observable from simulated data, then compute how that distribution deforms as the parameter of interest is varied, with adaptive binning and bootstrap resampling, 32 resamples throughout, to keep the estimate within error tolerance.¹

The subtle intellectual work is in the distinction the paper draws between two kinds of criticality. The FIM of the theoretical control parameter identifies the global operating regime: where the system sits on the axis from quiescent to epileptic, to put it biologically. The FIM of the observed branching ratio characterizes something different: the statistical sensitivity of the instantaneous dynamical state, and in particular how closely the current activity resembles self-sustaining critical balance. A peak at branching ratio 1 therefore appears as a temporal signature of critical-like balance, and how sharp and tall that peak is depends on how close the whole system is to its global critical point. One number tells you where you are; the shape of the peak tells you how sharp your sense of it is.

Two empirical findings give the mechanism texture. In the spiking network, the branching-ratio FIM peak in the supercritical regime barely changes as the control parameter moves, a behavior the authors attribute to inhibitory feedback bounding average activity regardless of how supercritical the connectivity is. And the whole-brain model, despite regional heterogeneity handled by weighting activity with the leading eigenvector of the structural connectome, reproduces the same peak signature, with robustness to threshold choice shown in the supplementary material.¹

Where a skeptic should push

The most load-bearing assumption is stationarity. FIM estimation assumes the probability distribution being estimated is the system's stationary distribution, and developmental or adaptive drift biases every downstream quantity. This is not a footnote for biological preparations; it is the central practical obstacle, and the authors name it themselves as the main barrier to extending the framework to experimental data.¹

Second, "model-free" is a qualified term. The method needs no knowledge of the true control parameter and no avalanche definition, which removes the two most criticized methodological choices in the criticality literature, where threshold and binning decisions famously swing inferred exponents.² But it introduces its own: the binarization threshold, the adaptive bin width, and estimator hyperparameters all remain free choices. The authors report robustness across these choices in the supplementary material, yet in a small culture with few channels, every distribution estimate is fragile, and no robustness sweep fully cures small-sample noise.

Third, the data budget is steep. The FIM is estimated from probability distributions of activity, which demands fine spatial resolution and long observation windows; the authors concede a full finite-size scaling analysis was too numerically demanding to include. And the heterogeneous-network problem is real: different regions can sit at different effective distances from their local critical points, producing extended near-critical regimes that a single global number will blur. The paper handles this by eigenvector weighting in the brain model, but in a 200-micrometer organoid with a few hundred electrodes, spatially resolved FIM is close to the noise floor.¹

Finally, the entire validation is synthetic. The models are well-chosen, spanning mean-field to connectome-constrained, but a criticality estimator that has never seen a single biological trace remains a promise. Demonstrated: the estimator locates critical points and tracks distance to them in three models, with sensible finite-size behavior. Asserted but unproven: that it survives contact with real neural recordings, where nonstationarity, sparse sampling, and unknown effective dimensionality all conspire against distribution estimation.

A control dial for organoid operating points

The authors write that their tool is promising for empirical recordings from multi-electrode arrays, calcium imaging, or fMRI. That sentence is, almost word for word, the organoid intelligence instrumentation problem: everyone in the field agrees cultures should be operated near criticality because that is where excitable networks show maximal dynamic range and information capacity, and nobody has a trustworthy, feedback-compatible gauge of where near actually is.³⁴

The non-obvious contribution is negative as much as positive. A great deal of organoid electrophysiology, knowingly or not, tunes preparations by making avalanches look right: threshold, bin, fit a power law, adjust the culture until the branching ratio hovers near 1. This paper shows the raw branching ratio statistics cannot locate criticality; the residence time dips rather than peaks at the transition, so a culture tuned until the ratio sits near 1 as often as possible may be tuned as far from criticality as its fluctuation statistics allow. The usable signal was never in the observable's mean behavior. It is in the observable's information geometry, which is a more demanding thing to compute but a far more honest thing to control against. For a field prone to curve-tuning, that correction has real value.

The opportunity is a closed-loop control variable. An FIM-of-the-branching-ratio readout is computable from spike counts, which is exactly what an array gives you, and it returns a continuous scalar, not a fitted exponent with confidence intervals. Peak width supplies tolerance bands; peak height supplies a sensitivity budget. A controller adjusting stimulation rate, pharmacology, or media conditions could hold a culture at a chosen distance from criticality the way a thermostat holds temperature, and experiments could then report an operating point, rather than a preparation history, as a reproducible experimental variable. As a maturation marker the same quantity is suggestive: a developing network's peak should sharpen as effective size grows, giving a single number that tracks functional maturation alongside the morphology everyone already images.

The threats are specific. The stationarity requirement collides head-on with developmental drift, which is not noise in a culture but the phenomenon itself; a controller stabilizing the FIM readout could be stabilizing the culture against its own development, and a preparation pinned at a fixed operating point is no longer the self-organizing system the field wants to compute with. The global-versus-instantaneous distinction matters even more under feedback: a controller holding the instantaneous branching-ratio FIM at its peak may be holding transient dynamics critical while the global regime slides subcritical, precisely the confusion the paper's two-tier analysis warns against. And the deepest caveat is that the functional premise itself is contested; whether criticality is even the right setpoint for brain function is an open argument, so an elegant dial for a contested setpoint should be adopted as a measuring instrument, not as a destination.⁴

The bottom line

In silico this is a careful, genuinely useful addition: a susceptibility-based estimator that needs neither avalanche definitions nor knowledge of the hidden control parameter, validated across three models with honest limitation sections. For organoid intelligence it is best read as the missing metrology layer between the slogan operate near criticality and the engineering reality of doing so. What would confirm it for tissue: FIM-of-the-branching-ratio tracking a culture through a pharmacologically induced, independently verified transition, with the peak following. What would break it: real array data where nonstationarity and sparse channels flatten the peak beyond recognition, in which case the field needs the estimator's robustness to be rebuilt for small, drifting systems before any dial gets attached.