The controller a survey astronomer already runs
A survey astronomer does not command one instrument. She commands a schedule of them: wide-field imagers throwing out alerts by the tens of thousands per night, transient brokers filtering that stream against known classes, spectrographs and follow-up telescopes allocated by hand or by heuristic, archival plates consulted when a candidate needs a decades-old baseline. The job is control, in the engineering sense — decide where to point, when, for how long — applied to a plant whose dynamics change with what is happening in the sky. That plant does not sit still long enough to be linearised once and forgotten. It has an operating point, and the operating point moves.
This is where gain scheduling, a technique with nothing to do with astronomy on its face, turns out to describe exactly the astronomer's problem. Gain scheduling is a method for controlling a system whose behaviour shifts with conditions: linearise at a set of chosen points, design a fixed controller at each, interpolate between them using a measured scheduling variable — Mach number for an aircraft, pitch angle for a turbine blade. For the astronomer, the scheduling variable might be a transient's rise time, its host-galaxy offset, its colour evolution. The controllers are follow-up strategies tuned for particular classes of event: this cadence for a nova, that spectroscopic urgency for a kilonova candidate, this archival cross-match for a possible AGN flare. The guarantees, as in any scheduled system, are local. They hold near the point they were tuned for. They say nothing about what happens between points, or about points nobody anticipated.
Where the schedule breaks
The characteristic failure is blunt: a transient fades before anyone allocates the telescope. This is not a rare edge case in time-domain astronomy; it is close to the median outcome for the interesting tail of events. A kilonova candidate brightens and reddens on a timescale of hours to days. A young supernova's early shock-cooling signature is visible for less than a day. An alert broker flags thousands of candidates a night, most of them junk, a handful of them the reason the survey exists — and the interval between the alert landing in a queue and a spectrograph being pointed at it is exactly the interval in which the scheduling variable, the thing that tells you which controller to apply, is itself changing state. By the time a human has classified the alert as worth the follow-up, the event has moved to a different operating point, or expired.
Shamma and Athans's result from around 1990 gives this failure a name rather than leaving it as folklore. Scheduled control is stable, they showed, when the scheduling variable moves slowly relative to the closed loop's own dynamics. The astronomer's closed loop — detect, classify, allocate, observe — has a characteristic time set by telescope scheduling overhead, human review, and instrument changeover; call it hours. Many of the events worth chasing have scheduling variables that move on the same timescale or faster. The condition for the schedule to be trustworthy is violated by construction, for a meaningful fraction of the survey's most valuable targets. This is not a defect in the astronomer's competence. It is what the mathematics says will happen when a slow loop is asked to track a fast variable.
Two positions, both defensible
Set against the failure is a genuine success story, and it deserves stating in full before it is qualified. Broker-based scheduling — assign gain by class, interpolate by measured feature, restrict effort to the well-characterised part of parameter space — is why time-domain surveys work at all. No survey team tries to write one universal follow-up policy for every possible transient; that would be as absurd as a flight controller with one gain for every Mach number. Instead the community has built, over roughly fifteen years of survey operations, an increasingly fine grid: photometric classifiers that place a candidate in one of several dozen recognised categories, each category carrying its own tuned response — spectroscopic priority, cadence, alert routing. Where the grid is dense and the categories genuinely recur, this works, and it works because it does not try to do more than it can certify. A superluminous supernova candidate gets a known response. A recurring nova in a catalogued system gets another. The schedule is legible, auditable, and does not surprise the observer with instability the way an uncontrolled adaptive scheme might.
Continuous re-evaluation of every alert against every incoming stream is not the safe option here — it is the dangerous one. Telescope time is the scarcest resource in the field. A system that keeps revising its priority queue as new photometry arrives will thrash: reallocate, cancel, reallocate again, and burn the very follow-up window it was meant to protect. Fixed classes with fixed responses are certifiable. That is not a limitation. It is the point.
This objection is not wrong about the danger. Adaptive scheduling that reacts to every noisy data point is a documented failure mode in control theory generally, not just in astronomy — parameter drift and oscillatory reallocation are real risks, not hypothetical ones. A telescope queue that revises itself every five minutes because a data point arrived is worse than one that commits to a plan and executes it. The case for scheduled, class-based response is not merely defensible; for the well-characterised part of the transient population, it is close to optimal.
Where the second position has to give ground
But the argument for scheduling assumed something the astronomer's actual sky does not grant: that the categories are known in advance and the grid can be made dense enough. That assumption holds for recurring novae and for the bulk of Type Ia supernovae, where decades of archival plates and thousands of prior events have populated the grid thickly. It does not hold for the object that does not fit any trained classifier — the fast blue optical transient with no established rise-time template, the gravitational-wave counterpart candidate with no precedent at all, the interstellar object passing through only once. These are exactly the cases where the scheduling variable is not merely fast but unenumerable: nobody knows in advance which axis of behaviour will matter. Gridding a two-axis flight envelope is engineering. Gridding an axis nobody has named yet is not possible, and treating the absence of a grid point as grounds to withhold follow-up is not caution. It is refusal to observe, dressed as discipline.
The resolution the domain forces is narrower than either side wants. It is not that continuous, ungated re-evaluation of every alert should replace class-based scheduling — the thrashing objection stands, and no serious survey operations team should abandon fixed response classes for the bulk of its traffic. It is that the interval between observation and classification should itself be treated as a monitored, revisable quantity rather than a fixed batch step. What a Large World Model does for this problem is retune within a scene — hold a coherent follow-up policy for the duration that a given transient's behaviour is being actively tracked, then have nothing to say once the scene, the observing window, ends. That is real progress over a policy frozen at survey-design time and never revisited. But it still drops the object between scenes: the moment tracking is not the moment classification. What the further move — treating every stream, transient broker output, spectroscopic queue, archival cross-match, as continuously open and provenance-tagged, without waiting for a scene boundary to trigger reconsideration — buys is not better classification. It buys an audit trail on the interval itself: a record of how long between alert and action, attached to every candidate, that a survey astronomer can actually use to ask whether the schedule is being outrun.
That number is a rate comparison, the same one Shamma and Athans wrote into control theory three decades ago. Astronomy inherited it whether or not any broker's documentation states it. Naming it does not solve the allocation problem. It does tell a survey team exactly where their grid is thin, and exactly which transients they are structurally destined to lose — not to bad luck, but to an interval that was never going to close in time.