Predictive internal states, measured instead of assumed

What the work claims

Frederick Hayes, an independent researcher, claims that learned recurrent and spiking agents trained on a compact foraging task develop internal dynamics with a specific cluster of properties: they predict later task outcomes above chance, retain behaviorally relevant information under compression, encode the agent's internal energy state even when explicit energy features are removed, and are at least partly causally involved in control, because perturbing them degrades behavior or prediction. The paper calls this pattern a computational analogue of predictive allostatic organization, borrowing the framing from Barrett and Miller's account of categorization as predictive, compressive, and tied to energetic regulation, while explicitly disclaiming any biological validation.

This is a primary computational study: one small task, many probes, a frozen evaluation protocol, and public code. It should be weighted as careful instrument-building rather than as evidence about brains or organoids.

How it works

The task is deliberately minimal. Each episode plays out on a 7 by 7 grid with one agent, three resources, and three threats, over at most 40 steps. The agent sees only a seven-component observation: noisy relative directions to the nearest resource and nearest threat, its own normalized energy, episode progress, and a binary contact sensor. Actions are the four moves, stay, and consume, where consume only succeeds while standing on a resource and restores energy. The headline outcome is strict: full-safe-efficient success requires completing the task, suffering fewer than three threat hits, and consuming efficiently, defined as at least 50% consume precision and no more than three consume attempts per success.

Against random and greedy heuristic baselines, the author trains a trace-augmented GRU policy and several spiking variants whose membrane state, threshold and gain modulation, and operating temperature are instrumented as recorded variables. The evaluation protocol is the paper's real strength: after pilot development, the benchmark was frozen at 20 training seeds, 20 held-out test base seeds, 50 episodes per seed, and 10 environment variants that drop or delay individual sensory cues, producing 1.8 million behavioral evaluation rows. Internal-state analyses then ask supervised probes (regularized logistic classifiers, evaluated with grouped validation by training seed and permutation-label baselines) whether recorded internal features predict later outcomes.

The key numbers: early internal dynamics predict later full-safe-efficient success well above permutation baseline. The strongest case is the default spiking configuration at ROC-AUC 0.802 against a permutation mean of 0.498; the trace-augmented GRU reaches 0.715 against 0.503; deltas over permutation across models range from +0.212 to +0.304. Low-energy state remains strongly decodable from internal dynamics even after explicit energy-related features are removed, and evaluation-time perturbations to temporal state, sensory inputs, operating conditions, and the allostatic mechanisms themselves measurably shift behavior and internal prediction. Exact future event prediction (a specific future contact, consume, or threat hit) is weaker but above baseline.¹

Where a skeptic should push

The single most load-bearing assumption is that decodability plus perturbation sensitivity amounts to organized internal computation rather than to task-trivial correlations. Linear probes on internal features will always find some signal in a trained network; the permutation baselines and grouped-by-seed validation are the right controls, and the author uses them, but the effect sizes that matter (roughly 0.2 to 0.3 AUC above permutation) are moderate, and the task is small enough that much of the predictable structure could live in the trace-augmented observation rather than in the recurrent core. The feature-family controls exist precisely to address this, and they show the signal is distributed across trace, policy-head, internal-dynamics, observation, and allostatic variables, which is honest, but also means no single family carries the phenomenon.

Second, this is a single-author, single-task study with no external replication. The protocol discipline (frozen benchmark, seed discipline, public code) is better than most multi-author work, but every quantitative claim lives in one 7 by 7 world. The energy variable is an engineered constraint and the paper says so plainly; it is not a metabolism model, so the allostasis framing should not be read as evidence that spiking networks regulate like organisms.

Third, the paper is admirably clear that decodability is not competence and competence is not causality. The perturbation experiments get partway to causality by showing degradation, but degrading a system and watching performance fall is a low bar; it does not identify which internal variables do the work. The paper treats this as a feature (it defines an evidentiary program) rather than overclaiming, and that restraint is the right call.

What organoid recordings would need to show

The non-obvious gift to organoid intelligence is methodological. The field routinely faces the question this paper formalizes: an MEA recording from a brain organoid shows structure, perhaps stimulus-locked responses, perhaps decodable features. Does the tissue carry predictive internal state, or are we reading correlations off a self-exciting blob? The paper hands over a concrete checklist. One: show that early activity predicts later task-relevant outcomes above a permutation baseline, with train and test rigorously separated by preparation (the organoid equivalent of grouped-by-seed validation). Two: show the predictive signal survives compression, because a real organizational property should live in a low-dimensional subspace rather than in noise. Three: show that a physiologically meaningful internal variable, the tissue's metabolic or homeostatic condition, is decodable from the dynamics even when the obvious readout channels are excluded. Four: perturb the candidate mechanism, pharmacologically or electrically, and show behavior or prediction degrades specifically, not globally.

That fourth criterion is the one the OI literature most often skips, and it is the difference between correlational decoding and a causal claim about computation. Several published demonstrations of learning or computation in organoids rest on before-and-after comparisons of firing statistics; very few include the probe, compression, and perturbation battery this paper runs on a toy gridworld. The bar for claiming that an organoid computes should be at least the bar this paper sets for a 49-cell artificial agent, and right now it often is not.

The threat is symmetric and worth stating plainly: if this evidentiary standard were applied retroactively, a nontrivial fraction of organoid-computing claims would downgrade from "demonstrated" to "suggested". That is uncomfortable but healthy; the field's credibility problem is not solved by lowering the bar to stay ahead of silicon. There is also an opportunity hiding in the paper's energy-constraint design. Organoid systems are profoundly energy- and homeostasis-limited: metabolic state gates everything the tissue does. A readout channel keyed to the tissue's internal condition, as the paper's low-energy decoding is, may be more informative for steering and training biological computers than spike-count metrics, and it is a channel that purely digital systems only approximate with engineered proxies.

The bottom line

Established within this study: learned agents in this specific task carry moderately decodable predictive information about their own future success, encode their energy condition in internal dynamics, and respond to targeted perturbation. Not established: anything about biological brains, organoids, or allostasis as a theory of either; the allostasis language is scaffolding for operational criteria, not a finding. The lasting contribution is the criteria themselves, plus a reproducibility standard (frozen protocol, seed discipline, public code) that organoid studies, with their notorious preparation-to-preparation variability, should envy and copy. What would confirm the broader program is the same battery showing graded, interpretable structure in a task where the agent or the tissue demonstrably relies on it; what would break it is a demonstration that all the decoded signal traces to engineered features of the observation, leaving the recurrent core inert.