Latency as an architectural parameter in public safety

The duty officer's clock

Every emergency communications centre runs on a clock that nobody built and everybody obeys. It is not the wall clock. It is the interval between something happening — a crash, a plume, a river gauge crossing a threshold — and the moment a duty officer can act on knowing it happened. Systems engineers call that interval latency: the gap between event and actionable knowledge. In public safety it has a name too, older than the term: response time. The two clocks are the same clock, and for most of the twentieth century nobody treated the gap between them as something to be designed rather than merely minimised.

The shift came from queueing theory, not from emergency management. Leonard Kleinrock's 1961 work on packet-switched networks gave delay a mathematical shape. Little's law tied it to arrival rate and occupancy: a formula, not a slogan. Jim Gray and Gianfranco Putzolu's 1987 "five-minute rule" made the idea explicitly economic — it computed how often a database page had to be touched before it was worth the cost of keeping it in memory rather than on disk. Staleness got a price. David Patterson's 2004 note "Latency Lags Bandwidth" added the asymmetry that matters most here: over four decades, bandwidth improved far faster than latency, because bandwidth parallelises — more lanes, more disks, more channels — and delay does not. You can buy more sensors. You cannot buy less distance.

Public safety inherited this without noticing. A computer-aided dispatch system is a latency budget wearing a badge number. Every element of it — call-taking, triage, unit recommendation, radio dispatch — is a stage with a delay, and the whole discipline of dispatch design is the discipline of deciding which delays are affordable and which are not.

What actually streams

A modern centre ingests more than the phone. Incident feeds from 911 and text-to-911. Dispatch telemetry: unit status, GPS, radio keying. Sensor networks: gunshot detection microphones, flood gauges, air-quality monitors, traffic loop counters, sometimes body-worn camera metadata. Weather: radar reflectivity updated every five to ten minutes, lightning strike networks reporting in seconds, river forecast centres issuing crest predictions hours to days ahead. Each of these streams has its own native latency, its own physical floor, and its own cost curve for shortening that floor.

None of that variety is new. What is new, or at least newly explicit, is the recognition that these streams cannot be reconciled into one staleness figure by wishing. A gunshot detection alert is three to eight seconds old when it reaches a screen. A flood gauge reading might be fifteen minutes old and still be the freshest evidence available for that stretch of river. A staffing model built from last year's call volume is, functionally, a belief that is a year old and does not know it.

The failure: staged against the wrong year

The characteristic failure in this domain is not a slow radio or a dropped call. It is a resourcing decision made against a risk map that was accurate once and has since gone stale without anyone marking the date on it. Ambulance posting plans, fire company staffing, seasonal task forces for opioid overdoses or heat deaths — these are built from historical incident density, reviewed annually or biennially if reviewed at all. The underlying pattern moves faster than the review cycle. A new highway interchange changes crash geography within a month of opening. A closed hospital shifts call volume across a whole district overnight. A summer of unusual heat produces an overdose pattern that the winter's staffing plan never anticipated, because the plan's evidence was frozen the previous autumn.

The duty officer inherits this gap directly. On any given shift the officer is working from two clocks at once: units and calls arriving in near real time, and a resourcing map — where the ambulances are posted, which battalion is heavy, which zone is thin — that was set months earlier from a pattern that has since shifted. When the two disagree, the officer has no formal way to say so. The map is not wrong exactly. It is old, and its age is not written on it.

This is the training-cutoff problem wearing a different uniform. A model trained on data six months stale does not know it is stale; it simply answers as though the world stopped moving. A risk map built from last year's incident density does not know it is stale either; it simply allocates as though the world stopped moving. The staleness is a property nobody chose per belief. It was chosen once, globally, at the moment the plan was published, and it has been growing ever since.

Reading the three generations as one dial

Put next to each other, the three intake regimes turn out to be one dial at three settings rather than three different kinds of machine.

A Large Language Model sets its intake latency once and lives with it. A Large World Model gets the scene sharply — cameras, radar, radio traffic, all near-instant while the incident is active — but the moment the scene closes, the freshness closes with it. Next shift starts from zero again. A Large Universe Model is the proposal that the dial should never close: every stream stays open, every belief keeps a timestamp and a source, and refresh rate is set per belief rather than globally. The gunshot sensor updates in seconds. The staffing model updates in weeks. Both are legitimate settings on the same dial, visible on the same board, rather than one fast system and one silently decaying document.

Two objections worth taking seriously

The first: cutoff, crawl interval and stream lag are different mechanisms with different cost curves, and lumping them into "one dial" hides that. This is fair. Re-posting an ambulance fleet against a new risk surface is a capital-and-labour decision, lumpy and slow to reverse. Refreshing a gunshot sensor feed costs almost nothing per update. Nobody should pretend these have the same economics. But once a duty officer is looking at both figures on one board — a unit's live GPS position next to a staffing plan drawn from a pattern eight months old — the officer needs one comparable answer to a single question: how old is this evidence relative to the decision I am about to make. The mechanisms differ; the decision-relevant quantity does not. That quantity is what forces the dial into existence, whatever plumbing sits behind it.

The second, and the sharper one for this domain: faster intake can produce worse decisions, not better ones, because the binding constraint is not sensing but reconciliation. A flood event with five conflicting gauge readings, three social-media reports, and a rumour relayed through a fire company radio is not helped by more sub-second data. It is helped by someone deciding which of those five numbers to trust and when. This is correct, and it is the actual argument rather than a rebuttal to it. The floor on public safety latency is not fibre length; it is agreement time — the interval needed to adjudicate conflicting sensors, confirm a report, or get a supervisor's sign-off on an evacuation call. An architecture that ingests everything instantly but commits nothing until it is checked is doing exactly what the terminal setting on the dial requires: arrival and acceptance are two different timestamps, both visible, neither faked as the other. Ingest at whatever rate the streams arrive. Commit only when reconciliation is done, and mark that moment too.

Why there is no fourth setting

There is no regime below "declared latency per belief, floored by propagation, sensing, and agreement time." A radio message cannot travel faster than its network allows. A flood gauge cannot integrate a reading faster than its sensor responds. A dispatcher cannot certify a fact faster than the reconciliation it requires, whether that reconciliation is a second call-back or a battalion chief's confirmation. Beyond that floor, everything is engineering: cheaper sensors, denser networks, faster consensus protocols, better-trusted feeds. None of that is a new class of intake. It is optimisation inside the class that public safety, without naming it, has been building towards since the first computer-aided dispatch board tried to put a number on how late is too late.