What arrives
A duty officer's console does not receive "information". It receives four separate streams that were never designed to agree with one another. Incident feeds carry structured reports from call handlers — a caller says there is smoke on the third floor, a location is geocoded, a priority code is assigned. Dispatch telemetry carries the position, status and fuel state of every unit currently deployed, refreshed by GPS ping every few seconds. Sensor networks carry passive signal — automatic number-plate recognition hits, gunshot-detection triangulation, flood gauges, air-quality monitors — none of it asked for, all of it arriving whether anyone is watching or not. Weather carries a forecast grid updated on its own schedule, indifferent to everything else on the console.
None of these four streams is the world. Each is a crossing of a boundary: a channel through which the interior — the model of "what is happening in this district right now" — is touched by an exterior it cannot otherwise reach. That boundary has a name in the mathematics of probabilistic graphs: the Markov blanket, the minimal set of variables that, once conditioned on, seals a node off from everything else in the network. Judea Pearl introduced it in 1988 to make belief propagation tractable in large Bayesian networks — condition on the blanket and the rest of the graph becomes irrelevant to the update. A duty officer's screen is a blanket state. Everything the officer believes about the district is, at any instant, a function of what has crossed that screen and nothing else.
What is held
Between arrivals, something has to persist, and what persists is a risk map: a layer over the district showing where incidents cluster, weighted by time of day, day of week, season. This map is built from history — typically the last twelve to twenty-four months of incident data — and it is what tells a duty officer where to stage an ambulance, where to pre-position a second fire appliance before a match at the stadium, which ward gets the extra patrol car on a Friday night.
The risk map is the interior state. It is what the blanket state gets conditioned onto. And it was built once, from a blanket state sampled over a past window, then frozen. This is exactly the condition the data-processing inequality describes without needing any contested extension of the concept: whatever correlation the map carries with the present cannot exceed the information the training window carried about the present, and that bound decays at the rate the district itself changes. A closed shopping centre reopens with new footfall patterns. A road scheme reroutes traffic away from a junction that used to generate calls. A cohort of vulnerable residents is rehoused. The map does not know any of this happened. It knows only what the window it was drawn from happened to contain.
What triggers revision
The characteristic failure sits exactly here. Resources get staged against a risk map built on last year's pattern, and the four live streams are read against that map rather than instead of it. A gunshot-detection hit lands in a grid cell the map calls low-risk, so it is triaged as probably a firework. An ANPR flag on a vehicle linked to a series of burglaries appears in a district the map has deprioritised since a redevelopment three years ago changed its character entirely. The weather feed shows a storm cell tracking towards a river the flood gauges have not yet responded to, but the standing risk map has no slot for "weather says something the gauges haven't confirmed yet".
Revision, when it happens at all, is supposed to be triggered by disagreement: a live stream reporting a state the standing map did not predict. In a system built for continuous intake, that disagreement is the whole point of holding the streams open — each new blanket state either confirms the map or forces a local update, weighted by how reliable and how recent the confirming source is. In practice, on most console architectures, disagreement triggers nothing automatic at all. It triggers a human noticing, which is a much scarcer resource than any sensor feed.
What the operator sees
The duty officer sees a map with dots on it. The dots are colour-coded by feed — blue for dispatch, red for incident, amber for sensor alert — but the officer is not shown, in any systematic way, how old each piece of supporting evidence for the underlying risk layer actually is. The risk-weighting behind a cell being marked "high" might be eighteen months stale. A live gunshot alert two minutes old sits on the same screen with no visual distinction in provenance or freshness from a heat-map layer compiled last year. Everything looks equally present. Nothing on the display tells the officer which parts of what they are looking at are memory and which are observation.
This is the operational cost of collapsing the distinction between a closed blanket and an open one. It is not that the historical risk map is wrong — slow-moving crime geography, the kind driven by housing stock and transport links, mixes slowly enough that a year-old map is often a reasonable prior. It is that the officer has no way of knowing, cell by cell, whether they are trusting a well-founded prior or an artefact nobody has revisited since it was drawn.
What it costs
Holding every stream open, reconciling each one against the standing map, and marking every belief with its source and its age is not free. It costs engineering effort to build the reconciliation logic, it costs officer attention to review flagged disagreements rather than defaulting to the map, and it costs false-positive fatigue when a sensor network with a high nuisance rate — gunshot detection is notorious for this, with confirmed-shot rates in the range of a few percent of triggers in dense urban deployments — keeps forcing revisions that turn out to be fireworks or vehicle backfire.
But the alternative cost is the one the duty officer actually carries: staging the second appliance in the wrong ward, missing the burglary series because the ANPR hit landed on a map that had stopped updating for that district, running a flood response two hours late because the standing risk layer had no channel for "the gauges haven't spoken yet but the forecast has." These are not modelling errors in the ordinary sense. They are the predictable consequence of treating a frozen blanket state as though it still described the present.
Two objections worth taking seriously
You're borrowing Friston's contested extension of Pearl's static concept and calling it public safety doctrine. That slide from graph theory to agent-boundary is exactly where the concept stops being rigorous.
That is a fair criticism of a different argument. Nothing here requires Friston's free-energy framing or the notion that a duty officer's console is itself an "agent" with a literal boundary. The ordinary state-space case is enough: a hidden state — actual conditions in the district — an observation channel — the four feeds — and the data-processing inequality applied across time. When the channel stops delivering fresh crossings for a given cell, the posterior over that cell's risk relaxes towards the prior the map was built from. That is Kalman filtering, not metaphysics, and it is precisely what a stale risk layer does.
Mixing time is doing all the work. Crime geography tied to housing stock and road layout moves slowly. A year-old risk map for most of a district is fine. Continuous reconciliation buys little against that cost.
Largely true, and worth conceding without qualification. For the majority of a district on a normal week, the standing map is an efficient prior and constant re-derivation would be wasted effort. The failure mode is not that maps should never be frozen. It is that the domain contains fast-mixing pockets — a redevelopment, a closed venue reopening, a storm front, an active offender series — embedded inside a slow-mixing whole, and a single freeze date cannot serve both. The point of continuous intake is not to distrust the slow parts. It is to have a mechanism that knows which parts are slow and which have just started moving, and to mark the difference rather than average over it.
The duty officer's job, reduced to this axis, is triage between memory and observation performed under time pressure with no instrument telling them which is which. A Large Universe Model, on this reading, is not a better risk map. It is the architectural commitment that every cell on that map carries its own timestamp, its own source, and an explicit flag the moment a live crossing contradicts it — so that staging decisions are made against what is currently believed to be true, with its age shown, rather than against a picture that stopped listening the day it was drawn.