A transient fades before anyone picks up the phone
A kilonova candidate appears in the alert stream at 03:14 UTC. By the time a telescope allocation committee meets, discusses target-of-opportunity priority, and someone with director's discretionary time actually points an eight-metre mirror at the right patch of sky, the object has faded by two magnitudes and the spectrum comes back as noise with a marginal continuum. This is not a rare failure. It is the modal failure of time-domain astronomy: the object was real, the alert was correct, and the loop between detection and action was too slow to close before the physics finished happening. Statistical process control, the century-old discipline of deciding from measurements taken as work proceeds whether a process is still behaving as it has been behaving, was built to solve exactly this shape of problem in a telephone factory. It generalises to a sky survey with fewer changes than you would expect.
What arrives
Nothing in modern astronomy waits to be collected into a corpus before it is examined. A wide-field survey camera images a patch of sky, differences it against a reference template, and emits an alert packet — position, magnitude, filter, a real/bogus score — within minutes. The Zwicky Transient Facility issues on the order of a million such alerts on an active night; the Vera Rubin Observatory's Legacy Survey of Space and Time is designed for roughly ten million a night once operating. Alert brokers such as ALeRCE and Lasair ingest that stream, cross-match it against known catalogues, and forward a filtered subset to scientists watching for particular classes of object. In parallel, spectroscopic follow-up requests arrive from telescopes that took the bait the night before, and slower streams arrive from archives — digitised photographic plates going back to the 1880s, in projects like DASCH, which let a present-day flare be checked against a light curve nobody thought to look at for a hundred years. Four intake channels, running at four different speeds, none of which stops.
What is held
The unit under observation is not the image. It is the object's evolving light curve, together with everything needed to interpret a new point on it: instrument, filter, seeing, air mass, the template used for differencing, the classifier version that produced the real/bogus score. A magnitude without that provenance is close to useless — a 0.3-magnitude jump might be the object, or it might be a change in seeing that widened the point-spread function and corrupted the photometric aperture. What the system holds, object by object, is a running estimate: current classification (nova, supernova subtype, tidal disruption event, spurious), expected fading rate given that classification, and a confidence that decays if no fresh data point arrives within the object's characteristic timescale. A fast blue optical transient loses its informational value in days; a slowly evolving supernova can go a fortnight between meaningful updates. The decay rate is itself part of what is held, because it determines how urgently the next observation is needed.
What triggers revision
The chart is the light curve plotted against the fading track expected for whatever the object was last classified as, with bands derived from the scatter of previously confirmed members of that class — not from statistical convention, but from the survey's own history of similar objects. A point landing inside the band changes nothing. A point landing outside it is an assignable signal: the object brightened when it should have faded, its colour shifted redward faster than the population, its position moved by more than expected astrometric noise. Any one of these reclassifies the object's urgency and, sometimes, its physical type outright. AT2018cow, one of the fast blue optical transients, was flagged for rapid follow-up precisely because its rise time sat far outside the band any known supernova occupies — the deviation was the discovery.
The Shewhart discipline that matters here is not the plotting. It is the restraint. A hundred simultaneous charts running at conventional three-sigma limits will, on average, throw a false alarm roughly once every four data points across the set — a fact that maps directly onto broker dashboards carrying thousands of live objects. Widening the limits to survive that noise buys silence at the cost of missing genuine events; narrowing them buys sensitivity at the cost of an alert queue nobody can work through. Real brokers set their thresholds the way Shewhart set his — on the economics of the response, not on a probability table — because a spectroscopic follow-up slot on an eight-metre telescope costs the observatory something whether or not the target is real.
What the operator sees
The survey astronomer on duty does not see raw pixels. She sees a vetting page: a light curve with control bands, a classification probability, a provenance stamp for every point, and a ranked queue of objects whose current state has just crossed a threshold worth acting on. The judgement she makes is the one Shewhart's chart was built to support — is this still behaving as expected, or has something changed that a human needs to decide about — compressed into an interface she can work through in the ninety seconds she has before the next alert arrives. What she cannot see, and does not need to, is the full stream underneath. The chart has already done the work of separating the object's ordinary photometric noise from the deviation that means the classification, or the urgency, has to be revised.
What it costs
Every escalation from "watch" to "point a telescope" spends something finite: minutes of an eight-metre instrument's night, a slot on a queue-scheduled spectrograph, the attention of a duty astronomer who can act on perhaps a few dozen genuine escalations in a shift, not the thousands the alert stream could in principle generate. This is where the funnel experiment's lesson bites hardest. Reacting to every deviation — chasing every alert as though it demands immediate follow-up — does not improve detection. It burns the telescope time that a real kilonova needed, on objects that a wider band would have shown were ordinary variable stars. The cost structure of astronomical follow-up is precisely the argument for computing limits from the survey's own history rather than treating every anomaly as equally actionable.
Two objections worth taking seriously
The sky is not a pressed washer. A control chart presumes a stable, repeatable process with a defined measurand. Transient astrophysics has neither: each object is a different physical process, observed once, under conditions that never repeat.
This is correct about individual objects and wrong about the population. No single kilonova recurs, but the class of kilonovae has a fading rate, a colour evolution, and a scatter around both, built from the sample observed since GW170817's optical counterpart AT2017gfo in 2017. The chart's baseline is not the object's own history — it has none — but the population's. That is a real weakening of the Shewhart architecture, and it means the bands are wider and less trustworthy for the first few members of any newly discovered class. The response has been to hold classification itself as a revisable belief, updated as the reference population grows, rather than pretending the measurand was ever fixed.
The bottleneck was never observational. Rubin's alert stream did not fail on the night AT2018cow-like objects were missed; the telescope allocation committee, the proposal review cycle, and the reluctance to interrupt a scheduled program for an unconfirmed transient failed. Deming's history says the same: charts existed for decades before anyone let a line worker act on one.
Accepted without much qualification. Director's discretionary time, target-of-opportunity interrupts, and rapid-response partnerships between survey and follow-up telescopes are institutional inventions, not statistical ones, and they are the actual determinant of whether a fading transient gets caught. The intake side of the problem — a continuously revised estimate of what an object is and how urgently it needs light collected — has a floor that current brokers already sit near. What remains above that floor is trust between institutions and authority to interrupt a queue, which is exactly the residual Deming's Japanese seminars addressed and Shewhart's American factories mostly did not. Conceding this is not a retreat from the claim; it is the claim, stated precisely. Observation terminates at continuous, provenanced, revisable belief. Everything above that is organisation.