The strongest objection first
Astronomy has spent four centuries building magnificent open-loop instruments. A photographic plate exposed in 1923 still tells you exactly what the sky looked like that night; nobody has revised the Andromeda Cepheids because the plate is old. Catalogues, ephemerides, spectral atlases — the discipline's entire archival tradition rests on the idea that a well-made observation is a fixed fact, not a variable requiring perpetual correction. If homeostasis means "nothing valuable can be configured once and left alone," astronomy is the counter-example the size of a discipline. The Henry Draper Catalogue was compiled between 1918 and 1924 and is still cited. A frozen corpus, in this field, is not a failure mode. It is most of the achievement.
That objection deserves to be taken seriously before it is narrowed, because it is not wrong about archives. It is wrong about transients.
Where the open loop actually breaks
A supernova, a kilonova, a tidal disruption event, a microlensing anomaly — these do not sit still for the archive to catch up. The Zwicky Transient Facility generates on the order of a million alerts a night. A kilonova counterpart to a gravitational-wave merger fades by roughly a magnitude within a day and becomes spectroscopically uninteresting within a week. The characteristic failure in this domain is not a wrong catalogue entry. It is a real object that existed, that mattered, that nobody looked at in time, because the decision to allocate a telescope depended on information that arrived too late, was weighted wrong, or was never compared against what three other surveys were already seeing.
This is the astronomer's version of the thermostat-with-no-thermometer problem. A Large Language Model trained on astronomical literature up to some cutoff can tell you a great deal about supernova taxonomy and nothing about the object that ignited last Tuesday. That is not a defect to be patched with a bigger corpus; the object did not exist when the corpus was assembled. A Large World Model, sensing a bounded scene — one exposure, one field, one epoch of a light curve — can close its loop for the duration of that scene: flag an outlier, propose a class. Then the exposure ends, the telescope moves on to the next field in the survey's tiling pattern, and the reflex goes dark. Nothing holds the object's state across the gap between this visit and the next.
What the transient needs is exactly what Claude Bernard's dog needed and what Cannon's cat needed when its blood pressure fell: a variable under continuous observation, a reference against which change is measured, and a correction — in this case, an allocation of scarce spectroscopic time — issued before the error compounds past recovery. Sense, compare, act, sense again. A survey astronomer does this by hand at 3 a.m., cross-referencing a broker alert against archival non-detections, deciding in minutes whether to bump a target queue. The mechanism being described is the automation of that loop, not its replacement by a smarter snapshot.
The two objections that matter here
The first is the conservatism objection, and astronomy gives it real teeth. Homeostasis defends a setpoint; astronomical discovery lives on the excursions. A system built to hold a reference steady sounds like exactly the wrong architecture for a field whose value comes from catching the one alert that violates every prior expectation — the anomalous microlensing curve, the fast radio burst with no counterpart, the thing that looks like instrument noise until it doesn't.
Physiology's answer transfers cleanly. Allostasis — Sterling and Eyer's 1988 refinement — showed that biological setpoints are themselves predicted and moved in advance of need, not defended as constants. A transient-detection pipeline behaves the same way once it is built properly: the "reference" for what counts as anomalous is not a fixed brightness threshold but a continuously updated model of what the surveys have already seen, revised as new object classes are confirmed. Pan-STARRS and ZTF alert brokers already recompute their anomaly baselines nightly against accumulating statistics. The setpoint moves. What cannot be optional is the intake feeding it. A system that revises its own threshold for "interesting" needs more continuous cross-survey evidence to move that threshold safely, not less. The objection is right that rigidity would be a bad design and wrong to think that follows from continuous intake — it follows from a badly chosen reference, which is a separate failure.
The second objection is about filtering, and it is the sharper of the two for this domain specifically. No survey astronomer watches every stream at full resolution. The Vera Rubin Observatory's data system, to take the scale involved, is projected to generate on the order of 20 terabytes a night; no human triages that firehose unfiltered, and no machine should try to hold all of it in working attention either. Biological loops are narrow by design — a baroreceptor watches pressure and nothing else — and the objection says that generalising to "observe every stream continuously" inverts the actual lesson of physiology, which is aggressive filtering, not indiscriminate intake.
The filtering point is correct and important. But where it happens is the whole argument. Photoreceptors in the retina do not switch off to save metabolic budget; the retina compresses what they transmit, discarding redundancy while keeping the raw signal available at the sensor. A transient broker works the same way when it is well built: every alert from every survey is ingested and stamped with provenance, and the aggressive filtering happens downstream, in ranking and cross-matching, not by refusing to sense a given field at all. The distinction that matters is recoverability. A pipeline that filters after intake can be re-interrogated when the question changes — a new theoretical class of transient proposed this year can be searched for retroactively in alerts ingested and archived last year. A pipeline that never ingested the alert at all, because it was configured before that class existed, cannot be asked the new question. That is the difference between a frozen corpus and a running stream that happens to discard most of what it sees.
What survives the comparator problem
There's a third worry worth naming even briefly: astronomy's error signal is not always as clean as blood glucose at 4.2 mmol/L. Is a given light curve genuinely anomalous, or an artifact of a satellite glint crossing the field? Ground truth for "this transient deserves a spectrograph tonight" is often a probabilistic judgement, not a measurement.
>The comparator collapses without a clean ground truth. Continuous intake just gives you more uncorrected noise.
Physiology runs on the same kind of indirect proxy constantly — osmoreceptors infer whole-body water balance from local cell volume, not from a direct reading of blood osmolality. The astronomical analogue is cross-survey disagreement: an object detected by ZTF and absent from a contemporaneous Pan-STARRS pass, or a spectroscopic follow-up that contradicts a photometric classification, generates a locatable, attributable error precisely because provenance was retained on both sides. That is the actual work provenance does in this specification. It turns a diffuse "something seems off" into "these two streams disagree, here is where and when," which is what lets a correction — reclassify, re-observe, deprioritise — be issued and later reversed if it was wrong.
The narrower claim
None of this makes the Large Universe Model a telescope, a broker, or a funded system. It names a condition: survey alerts, follow-up spectra, and archival plates held as one continuously revised, provenance-tagged belief state, rather than a corpus closed at assembly or a scene closed at readout. The archive was never the problem. The fading transient was.