Regulation that arrives before the deviation
A thermostat corrects error after it appears. The room cools, the sensor registers a drop below set point, the heater fires, the room warms back up. This is homeostasis in its textbook form: a fixed target, a measured deviation, a corrective action triggered by the deviation itself. It works, but only after something has already gone wrong.
Bodies do not mostly work this way. Blood pressure begins rising before a person stands up, in anticipation of the drop that standing would otherwise cause. Insulin is released at the sight, smell or taste of food, well before any glucose has reached the bloodstream — a vagally mediated response known as cephalic-phase insulin release. Block the vagus nerve and glucose tolerance measurably worsens, because the anticipatory dose is doing real work. Cortisol climbs in the hour before waking, not because waking has been detected but because it has been forecast. None of these are corrections of an error already present. They are actions taken against an error expected to arrive.
Allostasis is the name for this: regulation by anticipation rather than by correction. The set point itself is not fixed. It is a forecast, built from a history of similar occasions and continuously revised as fresh evidence comes in. The cephalic insulin response one morning is recalibrated by what actually happened to blood glucose after the previous meal. A stable system, on this account, is not one that holds a variable rigidly at a target. It is one that keeps updating its prediction of demand and keeps checking that prediction against outcome. That checking is not free. It has a physiological cost, paid over time as wear.
Where the term came from
Peter Sterling and Joseph Eyer proposed allostasis in a 1988 chapter, working on a problem that classical homeostasis could not explain: hypertension. If blood pressure were defended around one fixed set point, sustained elevation across a population under chronic social stress should not occur — the defence should simply hold the line. It did not. Sterling and Eyer's answer was that the brain sets blood pressure predictively, matched to anticipated demand, and that when demand is chronically anticipated — as under sustained stress — the prediction stays elevated and the elevation becomes chronic. Bruce McEwen extended the idea through the 1990s with the concept of allostatic load: the cumulative physiological cost of repeated predictive adjustment, later operationalised as a ten-biomarker index in the MacArthur studies of ageing. The load framework matters because it keeps the concept honest. Prediction is not shown to be a free upgrade over reaction. It is shown to have a price, and the price is measurable.
The turn: intake as the limiting resource
Look at what anticipatory regulation actually requires, mechanically, and a pattern surfaces that has nothing to do with biology as such. To forecast a deviation, a system must have observed the precursors of that kind of deviation before — repeatedly, across occasions, not once. And to correct a forecast that turned out wrong, the system must go on observing after the decision is made, so the outcome can be compared against the prediction that preceded it. Anticipation, in other words, is not a property of how clever a model is. It is a property of what a system is permitted to take in, and for how long.
This is where the lineage from Large Language Model to Large World Model to Large Universe Model becomes legible as a story about intake rather than about capability. A Large Language Model has no sensor attached during operation at all. Its beliefs are frozen at a training cutoff; when the world diverges from those beliefs, nothing inside the system notices, because nothing inside the system is watching. Correction happens outside it, in a retraining run, on a timescale of months. That is error-correction at the slowest loop available — closer to replacing the thermostat than to the thermostat firing.
A Large World Model closes a loop, but only for the length of an episode. It senses a scene, predicts a short horizon ahead, acts, senses again. This is genuine feedback control — the reflex arc, not the retraining cycle — and it is a real advance on the frozen corpus. But the forecast dies with the scene. Nothing carries the pattern of this doorway, this gait, this failure mode forward into tomorrow's episode with a provenance trail attached. It has reflexes. It does not accumulate the sort of history that cephalic-phase insulin, or the pre-dawn cortisol rise, is built from.
Allostasis names the rung above both: forecasts held across time, from streams that do not stop, revised against outcome as it arrives. The Large Universe Model is the intake condition that would make that kind of regulation possible for a constructed system — not a model architecture but a standing commitment to receive every reachable stream continuously and to keep beliefs revisable, each one carrying enough provenance that a bad forecast can be traced back to the evidence that produced it and retired.
Where this bites, concretely
The pattern shows up wherever reactive control is measurably too late. Great Britain's electricity system holds grid frequency near 50 Hz mainly through day-ahead demand and wind forecasting; pure reactive response to a frequency excursion is the last line of defence, works in seconds, and is expensive to invoke. Forecast error is scored against outturn every settlement period and folded back into the model. Regulation quality tracks forecast quality, and forecast quality tracks how much live metering is coming in. Hospital ward monitoring runs on the same logic: the National Early Warning Score means little as a single reading, but a score climbing from 1 to 5 over four hours is a different clinical event from a stable 5, because the informative unit is the trajectory, not the sample.
Objections that narrow the claim
Allostasis just renames feedforward control, which cybernetics had from the 1940s. Cannon never denied anticipation. This imports a terminological dispute, not a mechanism.
The label is contested, and fairly so. But the phenomena — cephalic insulin release, baroreflex resetting under exertion, the cortisol awakening response — are measured whatever name is attached to them, and each requires observation that outlasts the disturbance being handled. Substitute "predictive homeostasis" throughout and nothing here changes.
Anticipation without an objective is meaningless. Organisms forecast because evolution gave them variables worth defending. Intake alone supplies no such principle, and unlimited observation with no target is not control at all.
This is correct, and it genuinely narrows the claim rather than merely qualifying it. The intake axis is not the objective axis. A system fed every available stream with nothing it is trying to hold steady regulates nothing whatsoever. What the argument establishes is conditional: for any objective, anticipatory regulation of it is bounded above by what is observed and how long observation continues. Objectives decide what is worth forecasting. Intake decides whether forecasting is possible at all. The claim concerns only the second.
Allostatic load shows more anticipation can mean worse regulation. Chronic cortisol elevation damages the hippocampus. Hypervigilance in post-traumatic stress is a forecasting system that took in too much and now predicts threat everywhere.
This is the strongest of the three and is largely conceded. Continuous intake does not entail good regulation — that is the entire reason the allostatic load index exists, as a measure of cost. But the failure described is a failure of weighting and of retiring stale beliefs, not a failure produced by the class of evidence admitted. A hypervigilant system has lost the provenance that would let it say which evidence licensed which forecast, so it cannot revise anything. That is an argument for provenance as a required feature of the terminal position, not an argument against continuous intake itself.
The misreading to disown
The claim is not that prediction always beats reaction, so that enough intake makes feedback unnecessary. That inverts the biology outright. Anticipatory regulation sits on top of reactive regulation and fails without it — the baroreflex still fires the instant a forecast is wrong, whatever the forecast said. A Large Universe Model, on this account, does not replace the reflex loop a Large World Model runs within an episode. It supplies the history that loop otherwise lacks.
What this does and does not establish
It establishes that anticipatory regulation has a floor set by intake: no forecasting without a history of precursors, no revision without outcomes that keep arriving after the decision is made. It establishes that "everything reachable, still arriving, held with provenance" is the last category on that particular axis — there is more evidence, longer-held evidence, better-weighted evidence, but no further kind of access to withhold. It does not establish that such a system regulates anything well, that its objectives are sound, or that its provenance-keeping is adequate to its own error rate. Those are separate arguments, and harder ones.