The path, not the point
A body defends a temperature. A forest, in the season before it burns, defends something else: a course. Fuel moisture in a stand of Douglas fir does not sit at a value — it follows a curve that descends through spring, plateaus or dips sharply with a heatwave, and is expected to recover after rain. What matters to anyone managing fire risk is not "how dry is the litter layer right now" but "how dry should it be by now, and is it running ahead of schedule." That is a different kind of regulation from a thermostat, and it has a name.
C. H. Waddington coined it in 1957: homeorhesis, the stabilisation of a trajectory rather than a value. He was working on embryonic development, where a disturbed system does not snap back to a set point but resumes a developmental sequence — a chreod, he called it, a canalised valley the system falls back into after a shock. Homeostasis, Cannon's 1932 concept, could not explain this kind of robustness, because homeostasis has no notion of "on schedule." Homeorhesis does. Deviation is measured against where the system ought to be by now, not against a fixed level.
This distinction sounds academic until you ask what it costs to detect a deviation. A thermostat needs one number and a threshold. Something defending a trajectory needs the current measurement, a clock, an expectation for that clock reading, and enough history to tell a transient dip from a genuine departure. That is a much larger demand on intake, and it scales with the length of the trajectory being defended, not with the complexity of the system producing it.
Why this forces a ladder
Once you accept that trajectory control needs a position estimate, not just a reading, you can ask which kinds of system are even capable of computing one. A Large Language Model has read enormous numbers of trajectories — decades of fire history, thousands of fuel-moisture curves from a thousand watersheds — but it has no clock and no sense of where any of them stand today. Its corpus is a single frozen sample of the world's course, useful for describing shapes, useless for saying "we are twelve days ahead of the normal curing schedule this year." A Large World Model fixes the immediacy problem: it can sense a scene, hold a value steady within it, act on a live feed. But a scene has a beginning and an end. An episode that terminates supplies no elapsed time to measure a course against. You cannot ask a bounded scene how far along a season it is.
A Large Universe Model is the first arrangement on this axis where the residual — observed minus expected, at time t — is a computable quantity at all. That requires three things simultaneously: streams that do not stop, so there is always a "since when"; beliefs that are revised rather than overwritten, so a new sensor reading updates an estimate instead of replacing a snapshot; and provenance on every observation, so the system knows how much to trust a given input and when it was taken. Given those three, any trajectory question becomes answerable in principle. There is no further category of evidence such a system could want. What remains is longer records, better sensors, and earned trust in each stream — refinement, not a new kind of capacity.
Forestry and wildfire management is as good a proving ground for this claim as exists, because the discipline already runs on trajectories nobody can afford to observe once and be done with.
Where the claim is tested
An incident commander is the person responsible for this, in the sharpest possible sense. Fuel-moisture sensors report percentage moisture content in dead and live vegetation, updated hourly at automated weather stations, but sparsely across terrain. Satellite thermal passes — VIIRS at roughly 375-metre resolution, several overpasses a day if the orbit cooperates — give hotspot detections that are already stale by the time they're processed. Wind models run at mesoscale resolution and drift from observation within hours. Crew positions come from GPS trackers that report location, not exposure. None of these streams alone tells you where the fire behaviour sits on its expected trajectory for the day; together, timestamped and cross-referenced, they might.
The characteristic failure is brutally specific: an ignition is detected after the wind has already shifted, not before. A spot fire crosses a containment line during a wind change that was forecast six hours out but not confirmed on the ground until the fire behaviour analyst's model updated against a fresh observation. The commander was defending a trajectory — expected rate of spread, expected containment by nightfall — and the position estimate arrived late, because the stream that would have caught the shift, a ground-truthed wind observation near the flank, was thin exactly where and when it mattered.
This is a homeorhetic failure, not a homeostatic one. Nobody missed a threshold. Everybody missed being able to say, at 14:20, "we are currently running ahead of the expected spread curve for this fuel type and wind regime, and here is the confidence interval." The residual was uncomputable because the position estimate depended on a stream that had gone dark for the preceding ninety minutes.
The strongest objection, answered on its own ground
There is a genuine case that this whole line of reasoning overvalues continuous sensing. Spacecraft cross the solar system on an ephemeris and a handful of ranging passes because the dynamics are known and stationary; prediction covers the gaps cheaply, and continuous telemetry would be an expensive substitute for understanding. Fire behaviour analysts make the analogous argument constantly: fuel models, slope, and aspect are known well in advance, so a spread prediction run once at the start of shift should carry most of the load, with sparse checks against reality.
That argument holds wherever the disturbance is stationary and well characterised. It fails precisely where wildfire fails it: wind is not stationary, fuel curing accelerates non-linearly under heat, and spotting distance depends on ember lofting that no fuel model captures. Dead-reckoning error in a predictive fire spread model grows with the square of the unmodelled disturbance and the time since the last fix, and a wind shift is exactly the kind of disturbance that the model doesn't know about until an observation says so. Sparse sampling is optimal only when you already know the disturbance spectrum — and on a fire ground, that spectrum changes lot to lot, slope to slope, hour to hour. The ephemeris case earns its sparsity from decades of prior continuous tracking of orbital mechanics. Fire behaviour has no comparable settled physics at the scale of a single afternoon's wind.
A second objection cuts the other way, and deserves equal weight: that continuous sensing invites overreaction. Feedback theory is clear that high-bandwidth measurement paired with a naive controller chases noise, and there is a real version of this failure on a fireground — a commander who repositions crews every time a single sensor spikes, wearing out the response chasing artefacts rather than trend. This is a fair warning, but it is an argument about the control law, not about intake. A system can integrate observations continuously and still act on a filtered, hours-long window; bandwidth of sensing and bandwidth of response are different dials. The genuine concession is that continuous, provenance-tagged streams make spurious attribution easy — a false detection from a satellite pass mistaken for confirmed ignition — and that discipline in weighting sources by their provenance is exactly the hard part.
What is left to build
None of this argues that forestry agencies currently have anything resembling a Large Universe Model. What exists is a patchwork: hourly weather stations, several-times-daily satellite passes, wind models that drift, radios that report position but not exposure. The claim is narrower and more useful than a promise of readiness: the category of evidence needed to close the gap between "detected after the shift" and "detected before it" is not a new kind of sensor nobody has imagined. It is continuous, timestamped, trust-weighted streaming of the ones already named, held as revisable belief rather than a fresh snapshot each shift change.
That is the terminal rung on this axis. Beyond continuous, provenance-bearing, revisable intake, there is no further increment of observation that trajectory control could ask for — only longer records, better calibrated sensors on the flank where the fire actually is, and commanders trained to trust a residual over a hunch.