A rate set by the signal, not the watcher
Harry Nyquist, working on telegraph channels at Bell Labs in 1928, established a relation that later bore his name properly once Vladimir Kotelnikov proved it in 1933 and Claude Shannon gave the canonical statement in 1948–49: a signal containing no frequencies above some bound B can be reconstructed exactly from samples taken at any rate above 2B. Sample slower than that and the higher frequencies do not disappear. They fold down and reappear disguised as slow, low-frequency content. This is aliasing, and its defining cruelty is that the reconstruction built from aliased samples is smooth, internally consistent and wrong, with nothing in the sampled data itself flagging the error. The remedy control engineering settled on by the 1950s and 60s is the anti-aliasing filter: strip out what the sampling rate cannot resolve before you sample, so that ignorance is at least known, rather than quietly rendered as false confidence.
The threshold is a property of the thing being watched, not of the watcher's diligence or intentions. A duty officer who checks a feed twice as often has not solved an aliasing problem if the underlying process moves four times as fast as that. Nyquist converts a soft question — how often should I look? — into a hard constraint set by the bandwidth of the source.
From sampling rate to a lineage claim
Treat intake, generation over generation, as a sampling policy and the boundary is exact rather than rhetorical. A Large Language Model samples the world once, at a training cutoff. Its rate on any process that continues after that cutoff is zero, so every such process is fully aliased: the model reconstructs a static, plausible world from a single frame and reports it with total fluency, because nothing inside a frozen corpus can tell it that the frame has gone stale.
A Large World Model samples fast but only while a scene is present — camera and lidar rates adequate for a manipulation task running at 30 Hz, and then nothing at all once the episode ends. It resolves the grasp and is structurally blind to the corrosion that will compromise the same mechanism eight months later, because that frequency was never inside the window it was watching.
A Large Universe Model is the arrangement in which sampling rate stops being inherited from how the system was built and becomes a per-stream, revisable parameter, with provenance recording what rate produced each belief — and therefore which frequencies that belief could never have resolved. That is the anti-aliasing filter made explicit and permanent rather than applied once at design time. Below that arrangement, confident false reconstruction is not a risk to be managed. It is a theorem: undersample a signal and the discarded information cannot be recovered downstream, however good the model that consumes the samples.
What a duty officer actually watches
Public safety is a live test of this because a duty officer's desk is nothing but converging streams at wildly different natural bandwidths. Incident feeds arrive as discrete events — a call, a report — irregular by nature, sometimes seconds apart during a mass event, sometimes hours apart on a quiet shift. Dispatch telemetry updates unit position and status on the order of seconds. Sensor networks — gunshot detection, traffic loop counters, flood gauges, air quality monitors — sample continuously at rates fixed by the hardware that was procured, often years before the officer on shift ever saw the console. Weather arrives as a mix: radar refreshed every five to ten minutes, ground stations on longer cycles, forecast models updated a few times a day.
None of these rates were chosen against the bandwidth of what they are meant to catch. They were chosen against budget, bandwidth in the network sense, and the state of sensor technology at procurement. A flash flood gauge sampling every fifteen minutes is adequate for a river that rises over hours and useless for a flash flood that peaks in twenty minutes — the hydrological equivalent of filming a rotor blade at 24 frames a second and getting a sharp, stationary, wrong answer.
The characteristic failure: resourcing against last year's shape
The specific way this goes wrong in public safety is not random error. It is systematic: resources get staged against a risk map built from historical incident density, itself compiled from a fixed, coarse sampling of past events, and that map is then treated as a stable description of where the current shift's danger lives. The map is a low-pass reconstruction of last year's pattern. It is smooth. It is plausible. Nothing in the map itself indicates which of this year's higher-frequency changes — a new venue opening, a gang territory shift, a heatwave driving a spike in a particular ward — have aliased into an apparent continuation of an old trend.
A duty officer allocating patrol cars against that map is doing exactly what a control system does when it samples an unfiltered signal below Nyquist and trusts the result. The failure is invisible until the mismatch is large enough to produce an incident the resourcing did not anticipate, and even then the map does not confess which of its other conclusions were also built on aliased history. That is the theorem, not a metaphor: once the higher-frequency structure was discarded at the point the historical risk map was compiled, it cannot be recovered by staffing harder or checking the map more often. The information is gone; only re-sampling the live streams at a rate matched to how fast the city's risk actually moves can restore it.
Two objections a duty officer would raise
"We can't watch everything continuously and we don't need to. Most shifts run fine on scheduled patrol reviews and a morning briefing. Demanding continuous intake on every stream is asking for infrastructure no service can fund."
This is correct about the demand, and the argument does not actually make it. Decisions have their own bandwidth requirements, usually far below the sampling rate of the raw sensors feeding them. A staffing decision made once per shift does not need gunshot-detector-grade timing; it needs an estimate whose refresh matches how fast the underlying risk genuinely changes, which for staffing is plausibly hours, not seconds. Oversampling here would import noise — chasing single incidents as if they were trend shifts — at real cost. The actual demand is narrower and harder to dismiss: choose the rate against the process, not against the org chart, and revise it when the process changes. A gauge network built for slow river rises still needs a faster tier for flash-flood-prone catchments, chosen because that catchment's hydrology is fast, not because more data is always better. Public safety agencies that moved to five-minute severe-weather nowcasting after missing fast-forming convective cells did exactly this: they did not sample everything faster, they identified which stream had outrun its instrument and re-tiered that one.
"Event-triggered dispatch already does this — units are alerted on threshold crossings, not on a fixed clock. So the problem is solved without needing some universal continuous-intake architecture."
Event-triggered sampling is a real and valid refinement, and it is exactly the compressed-sensing move: buy economy by trading a fixed rate for a threshold rule. But the threshold rule has to be watched continuously by something for the trigger to fire at all — a sensor idle between triggers is still a sensor whose own bandwidth was fixed at procurement, and if the threshold was calibrated against last year's incident pattern, it can silently fail to fire on this year's faster or differently shaped event. Threshold-based dispatch relocates the aliasing risk from the response channel to the calibration of the trigger; it does not remove it. The trigger itself needs the same discipline the argument asks for everywhere else: a stated bandwidth assumption, a record of when that assumption was last checked against reality, and a mechanism for revising it when the ground truth — new venue, new season, new offender pattern — moves faster than the trigger was built to notice.
Provenance as the only durable fix
None of this is solved by adding more sensors or more staff hours in the abstract. It is solved, to the extent it can be, by making the sampling rate on each stream an explicit, recorded, revisable choice rather than an inherited default, and by attaching to every belief the resolution limit of the stream that produced it — so that when the flood gauge says the river is stable, the record also says at what rate it was watched and therefore what it could not have seen. That is the shift from a Large World Model's episodic competence to a Large Universe Model's standing discipline: not more data, but data whose limits are stated rather than assumed away. A duty officer cannot outrun a signal that moves faster than the instrument watching it. What can change is whether anyone downstream is told, in the moment, which instrument that was.