The clock that starts at first light
A survey astronomer's working day is bounded by a number most colleagues outside the field never think about: the interval between an exposure and the report that something in it has changed. The Zwicky Transient Facility scans roughly 3,750 square degrees a night. The Vera Rubin Observatory's Legacy Survey of Space and Time will issue on the order of ten million alerts a night once operational. Each alert is a claim that the sky, at one pixel, is no longer what the archive says it was. From the instant the shutter closes, the correlation between "what the plate shows" and "what is actually happening at that coordinate" starts to fall. It falls at a rate set by the physics of the object, not by the quality of the pipeline that found it.
This is mutual information decay, and astronomy is one of the cleanest domains in which to watch it happen, because the discipline has spent two centuries building instruments whose entire purpose is to reset a clock that never stops running.
Two positions, honestly stated
Position one: the sky is, for most purposes, a frozen corpus, and rightly so. Stellar positions, spectral classifications, the Hertzsprung-Russell diagram, the physics of stellar evolution, the calibration of standard candles — these are close to stationary on human timescales. Gaia's astrometric solution, once published, is not falling out of correlation with the galaxy at any rate an observing proposal needs to worry about. A Large Language Model trained on the astronomical literature would retain nearly all of its useful bits about these invariants indefinitely. Treating every observation as perishable is a category error that would paralyse the field: catalogues work precisely because most of the sky does not do anything.
Position two: the operationally decisive part of astronomy is exactly the part that is not stationary, and on that part the frozen-corpus stance fails within hours. A kilonova associated with a neutron-star merger is bright enough to trigger a follow-up campaign for perhaps a week and spectroscopically informative for a fraction of that. A superluminous supernova can be followed for months, but the early-time colours that discriminate progenitor models are gone within days. A tidal disruption event's X-ray flare can rise and fall inside an observing night. None of these obey the calendar of a journal publication cycle, a proposal deadline, or a telescope time allocation committee that meets quarterly.
Both positions are defensible because both are describing real parts of the same sky. The disagreement is not about whether decay happens. It is about how much of astronomical practice lives in the decaying fraction, and what follows from that.
What survives the freeze
Take the first position at its strongest. The Sloan Digital Sky Survey's photometric catalogue, released in stages between 2000 and the 2020s, remains scientifically load-bearing for population statistics, luminosity functions, large-scale structure — all quantities whose mutual information with the present sky decays so slowly that "present" barely applies. A quasar's redshift, once measured with a decent spectrum, is not going to un-measure itself. The asymptote of mutual information decay, the invariant residue that a snapshot retains forever, is large in astronomy, arguably larger than in almost any other observational science, because so much of the subject concerns objects whose relevant timescale is millions or billions of years.
This is the correct core of the "decay is overstated" objection, and it should be conceded without hedging. A frozen archive is not a lesser tool. It is the right tool for the stationary manifold. The mistake is inferring from this that the archive is therefore adequate for the non-stationary manifold too — the transient sky, the variable sky, the sky of things currently changing brightness, position, or spectral state. Those are precisely the objects on which the survey astronomer's actual professional anxiety concentrates, and they are a small fraction of all astronomical objects but a large fraction of the discoveries that matter for physics currently in dispute: kilonovae for the equation of state of nuclear matter, tidal disruption events for black hole demographics, fast radio bursts for the intergalactic medium. The invariant sky is the background. The decaying sky is where the argument is being had.
The cost of chasing it
Telescope time is the scarcest resource in the field. Spectroscopic follow-up on every alert would burn a facility's entire allocation on false positives within a week. Selectivity is not a failure of ambition; it is the only economically sane policy.
This objection is close to unanswerable as an engineering constraint, and it should be treated as the strongest challenge to any argument for continuous intake. The numbers make the case. A wide-field survey can generate an alert stream in the thousands per night; a given 8-metre-class spectrograph can realistically follow up perhaps a dozen targets in that time, fewer if each requires a long integration. The gap between what is flagged and what is confirmed is where transients are lost — not because no one saw them, but because no one could allocate the instrument before the light faded. A superluminous event caught at discovery magnitude and never spectroscopically classified is a permanently incomplete record: the alert exists, the photometric light curve exists, and the one measurement that would have fixed its physical type does not, because it was never taken while there was still signal to take.
The resolution is not "spend more" — no facility budget grows fast enough to keep pace with survey depth doubling every few years. The resolution is that mutual information decay, once it is decomposed per object class rather than treated as a single number, tells you exactly where to spend. A tidal disruption event's X-ray decay constant is different from a Type Ia supernova's optical decline, which is different again from a repeating fast radio burst's activity window. Brokers such as those built on the Zwicky and Rubin alert streams already rank candidates by predicted decay rate combined with scientific value, which is a crude but real implementation of per-variable refresh: observe fastest where the half-life is shortest and the payoff highest, not everywhere at once. Still, Sivak, Bell and Crooks's 2012 result — that retained information failing to predict the future is exactly the fraction paid for as dissipated work — has a direct translation here. Telescope time spent confirming a transient whose type could have been inferred from photometry alone is time paying for bits that were not going to help. The cost objection is right about the budget and wrong about the conclusion: it argues for triage, not for abstention.
Provenance as the instrument log
None of this triage is possible without knowing, for each belief in the archive, when it was last actually observed and by what. An entry in a transient broker that reads "classified: Type Ia, spectrum taken, epoch +3 days" carries a different half-life from one that reads "candidate: photometric colour only, epoch +11 days, no spectrum." The second is close to its expiry; the object may already have faded past the point where any instrument on Earth can recover its type. Treating both entries as equally "known" is the single most common way astronomical decision-making silently degrades — not through bad data, but through good data whose timestamp has been forgotten in the interface.
This is where the Large Universe Model framing earns its keep as an argued category rather than a slogan. It does not propose that decay be abolished — Lorenz's error-growth argument from 1963 has no astronomical exemption, and a transient's brightness is no more predictable indefinitely than tomorrow's weather. It proposes that decay be made computable, belief by belief, by carrying provenance through the pipeline: instrument, epoch, and calibration lineage attached to every classification, not just to the raw photometry. A survey astronomer with that structure in hand is not asking "is my archive fresh." They are asking "which of these four hundred alerts has a half-life measured in hours, and which in months," and allocating the one scarce spectrograph accordingly.
Where the resolution actually lands
Neither position wins outright, and the honest resolution narrows rather than settles the claim. The stationary sky is real, large, and correctly served by archives that behave like frozen corpora — nothing about continuous intake improves a redshift already measured. The non-stationary sky is real, comparatively small, and is exactly where the discipline's open questions live, which is why it dominates anxiety disproportionate to its share of the object count. Continuous, provenance-tagged observation does not defeat the decay of a fading transient; Lorenz's horizon and the telescope's finite aperture both still apply. What it changes is whether the system finds out it is losing the transient while there is still time to point an instrument at it, rather than after. That is a narrower claim than "observe everything, always." It is also the only version of the claim that survives contact with a fixed number of nights and a finite number of mirrors.