What arrives
A survey camera exposes the sky in tiles, several times a night, across a cadence designed to catch anything that changes. Zwicky Transient Facility does this at roughly 3,750 square degrees per hour; the Vera Rubin Observatory's alert stream, once at full operation, is expected to issue on the order of ten million alerts each night, one packet per source that brightened, faded, or moved relative to a reference image. Each alert is a difference: a point of light where the archive says there should be nothing, or a changed magnitude where there should be a known, steady star. Alongside the imaging stream there are spectroscopic queues, cross-matches against catalogues of known variables, and — for anything that might be historically informative — a search of digitised archival plates, such as the Harvard collection scanned under the DASCH project, holding photographic exposures back to the 1880s.
None of this arrives once. It arrives continuously, in packets that a broker such as ANTARES, ALeRCE, Lasair or Fink must triage in seconds, because the next tile is already exposing. A survey astronomer on duty is not handed a dataset. They are handed a rate.
What is held
The system cannot hold every alert as a live concern; it holds a ranked set of candidates, each carrying a belief rather than a fact. A candidate transient has a classification probability — supernova, cataclysmic variable, asteroid, artefact — attached to a light curve that is itself incomplete, since only a handful of epochs exist so far. It has a sky position with an uncertainty ellipse, a host-galaxy association with its own redshift error, and a provenance trail: which telescope, which filter, which pipeline version produced the photometry point that moved the classifier's confidence from 40 to 70 per cent.
This held state is small compared with the intake. Rubin's ten million nightly alerts collapse, after filtering, to perhaps a few hundred objects worth a human glance and a handful worth telescope time. What is kept is not the flood but a compressed, timestamped, sourced belief about each surviving candidate — exactly the object an anytime system is supposed to maintain: available at any instant, with a stated quality, ready to be revised.
What triggers revision
Revision is triggered by the next observation, and only by that. A second imaging epoch tightens the light-curve fit and can flip a classification — a candidate kilonova can turn out to be a slow nova as the colour evolves. A spectroscopic follow-up, if telescope time was allocated, replaces a photometric guess with a redshift and a line-identified type, collapsing the uncertainty by an order of magnitude in one exposure. A cross-match against an archival plate can demote a "new" transient to a previously catalogued variable star, discovered once in 1953 and forgotten. Each of these events updates the belief state for one candidate without touching the others; there is no global recomputation, because the other several hundred candidates have not received new evidence and their beliefs remain exactly as good, or as stale, as they were.
The trigger that matters most, though, is time itself passing without new evidence. A transient's error bars on type and rate of decline widen the longer it goes unobserved, even with no new data at all, because the model of "how fast can this class of object plausibly fade" is doing the work that a real photometric point would otherwise do. Silence is itself informative and must be represented, not ignored.
What the operator sees
The duty astronomer's console does not show a finished discovery. It shows a candidate list with, per row: current best classification and probability, magnitude and its trend, time since last observation, and an estimate of remaining visibility — how long before the object sets, or drops below spectroscopic reach, or simply fades past the limiting magnitude of every instrument that could still reach it. That last column is the deadline.
This is the anytime interface made literal. Nothing here is a final verdict. Everything here is interruptible: the astronomer can act on the ranking at minute four of the night or minute forty, and either way gets a defensible, quality-annotated answer, because the profile — confidence rising as spectra and later photometry accumulate, confidence eroding as visibility windows close — is displayed alongside the number, not buried behind it.
What it costs
The cost is allocation, and it is unforgiving. Spectroscopic time on a 4-metre-class instrument is booked in minutes that do not exist twice; a director's-discretionary-time request competes against every other duty astronomer's request that same night. Compute cost is smaller but real: cross-matching a candidate against a century of archival plates, or refitting a light curve against a library of template classes, takes seconds per object but the object count is in the thousands nightly, so the triage pipeline itself has a time budget it must respect or it becomes the bottleneck it was built to avoid.
The characteristic failure of this domain is specific and well documented: the transient fades before anyone allocates the telescope. A fast blue optical transient can rise and decline within days; a kilonova counterpart to a gravitational-wave event, such as AT2017gfo following GW170817, was spectroscopically followed only because an extraordinary, coordinated effort compressed the normal allocation delay from days to hours. Most candidates get no such effort. The cost of continuing to refine is measured against the cost of the window closing, and the astronomer's real decision is not "what is this object" but "is it worth telescope time before it answers that question for us by disappearing."
The profile that survives, and the one that does not
Anytime algorithms depend on a performance profile: quality improving predictably with more time spent on a fixed problem. A transient's classification is not a fixed problem — the target itself is fading, brightening, or evolving in colour while you compute. There is no stationary optimum to converge on, so there is no meaningful suboptimality bound, and calling the triage pipeline "anytime" borrows the term's authority without its guarantee.
This is correct about the mathematics and worth conceding fully. A convergence bound of the kind iterative deepening enjoys in a static chess position has no analogue when the object under study is itself non-stationary. What survives is weaker but still auditable: calibration, meaning whether a broker's stated 80 per cent supernova-probability candidates turn out to be supernovae roughly 80 per cent of the time when checked against later spectroscopic confirmation; and freshness, meaning the pipeline reports the age of the last photometric point behind every classification, so a stale 80 per cent and a fresh 80 per cent are visibly different things. Brokers that report neither of those are not doing anytime triage. They are doing fast guessing with a probability sticker attached.
An answer that's always on the console is always available to be acted on. A duty astronomer under time pressure will allocate the telescope on whatever number is showing, not on the curve behind it. Batch procedures that stay silent until a spectrum is in hand impose a discipline that a perpetually updating candidate list quietly erodes.
The concern is real, and the domain has paid for it: early photometric classifications have driven follow-up requests that a subsequent spectrum overturned, wasting an allocation that another candidate needed more. But the discipline being praised here is a crude stand-in for an explicit decision rule, and astronomy already has better versions of that rule: allocate when the expected scientific value of confirming this candidate now exceeds the expected value of waiting for one more epoch, discounted by the probability the object is gone by then. That calculation needs the profile — confidence, freshness, and remaining visibility — not its absence. A console showing a bare number with no history is a bad interface. A console refusing to show anything until certainty arrives simply guarantees the object fades unobserved, which is its own decision, just an unexamined one.
Why the loop has no stopping point
The survey does not pause between tiles for the pipeline to finish thinking, and the sky does not pause for the survey. There is no moment at which every candidate's classification is final, because finality would require every stream — imaging, spectroscopy, archival cross-match — to stop, and none of them will. The only coherent output at any instant is the one the console already gives: a current best belief per candidate, timestamped, sourced, and bounded by an honest statement of how much of the picture is still missing. That is not a lesser answer than a finished catalogue entry. Given what is actually running, it is the only kind of answer available, and the discipline lies in never pretending otherwise.