The strongest objection first
Here is the challenge stated as sharply as it deserves to be stated.
Thermodynamic entropy is a property of microstates under a Hamiltonian. A risk map is not a physical system with a partition function. There is no theorem that licenses moving the second law from gas cylinders to duty rosters. Calling a stale deployment plan "entropic" borrows the authority of physics for a claim physics never made.
This is correct as far as it goes, and it should be conceded without qualification. Nothing in Clausius or Boltzmann forces a risk assessment to degrade. There is no conservation law for the freshness of a call-volume forecast. If the argument depended on the literal apparatus of statistical mechanics, it would fail on contact.
But the load-bearing claim was never the entropy theorem. It is a weaker and much better-established result from Shannon: mutual information between two systems cannot increase without a channel connecting them. A control room's risk map and the city it describes are two systems. While the channel between them is open — incident feeds, dispatch telemetry, sensor networks, weather data all flowing in — the map can track the city. Close the channel, freeze the map at last year's pattern, and the city keeps changing while the map cannot. Mutual information between map and city then falls, not because entropy has some mystical grip on paperwork, but because information about a moving target cannot grow across a closed channel by definition. The physics is illustrative dressing. The information theory is the argument. It survives the objection intact, and it is narrower than the language of "entropy in public safety" suggests — which is the honest price of conceding point one.
What a frozen risk map actually is
Call the sealed case the Large Language Model position, translated into this domain: a risk assessment compiled from last year's incident data, printed, distributed to shift commanders, and treated as ground truth until the next annual review. The document does not decay in any physical sense. The percentages, the heat-map overlays, the staffing ratios recommended for each ward — all of it is bit-identical the day before the review as the day after it was issued. What has moved is the city. A retail unit that burned three times last winter has since been demolished. A road that flooded twice has had a culvert replaced. A housing estate flagged for anti-social behaviour has had half its population turned over. None of this shows up in the document, because the document has no channel back to the world it describes.
The characteristic failure this produces is precise and recognisable to anyone who has staffed a control room: resources are staged against a risk map built on last year's pattern, and the duty officer inherits both the map and the blame when it is wrong. A patrol car is pre-positioned near last winter's hotspot while this winter's hotspot, three streets over, has no coverage. Nobody involved made an error in the ordinary sense. The map was accurate when it was made. The failure is structural — a closed channel, applied to a moving referent — and it is asymmetric in blame, since the document did not fail so much as the world moved without it.
The stationary majority, and why it doesn't save the map
The second objection worth taking seriously is that most of what a risk map encodes is not volatile at all.
Road layouts, building heights, hydrant locations and population density change on the scale of years or decades. If ninety per cent of the underlying geography is stationary, an annual refresh is a modest tax on an otherwise durable asset, and a live feed is overkill for information that barely moves.
This is true, and it explains why annually reviewed risk maps remain useful rather than worthless. Nobody proposes re-surveying building heights hourly. But the value of a fact to a duty officer is not proportional to how slowly it changes. It is proportional to how much a wrong answer costs at the moment of decision, and the decisive facts in public safety cluster heavily in the volatile minority: a road closed for gas works since Tuesday, a football fixture pulling eleven thousand people through a junction that the annual map treats as ordinary, a flood warning issued forty minutes ago, a firearms incident three streets from where a patrol was staged against last year's burglary cluster. Stationary geography sets the backdrop. Volatile events decide whether resources arrive on time.
Attaching a live incident feed or a weather overlay to the annual map is often proposed as the fix, and it is a real improvement — but it concedes the argument rather than resolving it. It is an admission that a channel to the present must stay open for the plan to mean anything operationally. Once that is granted, the only remaining questions are how many streams, how continuous, and with what accounting for reliability. That is not a patch on the closed model. It is a move to a different position on the intake axis altogether.
The bounded case, and its honest limit
Between the sealed annual map and continuous multi-stream ingestion sits an intermediate position worth naming properly: the sensor feed that is live but narrow — a single CCTV bank, a single flood gauge, a single weather station, attended only while someone is watching it. This is the Large World Model position translated into control-room terms. It genuinely tracks the present, but only for the scene in front of it, and only while attention is paid. A camera on a known junction will correctly show a live crash. It will say nothing about the ambulance stuck on the ring road four miles away, because no channel runs there. The correspondence is real but local, intermittent, and silent about everything outside its bounded scene.
| Position | Intake | What it knows about the present | Characteristic failure |
|---|---|---|---|
| Sealed risk map | Closed after annual review | Nothing beyond compilation date | Staged against last year's pattern |
| Single live feed | Open but narrow, attended | Only the watched scene, only while watched | Blind outside the frame |
| Full stream ingestion | Open across incidents, dispatch, sensors, weather | Current across everything still reporting | Requires provenance to filter noise |
The contamination objection, and what actually answers it
The most serious objection remaining is not about definitions. It is about hazard.
Opening every channel means ingesting rumour with fact: a social media report of shots fired that is actually fireworks, a sensor drifting out of calibration, a feedback loop where a control room's own broadcast becomes the next unverified report it receives. A sealed map cannot be poisoned. It can only be old.
This is largely right, and it is the strongest case against continuous intake as such, not merely against its cost. The answer is not to close the channel back down — a fridge does not achieve cold by sealing its door, it achieves cold by continuously pumping heat out against a filter on what enters — but to make provenance and decay part of the ingestion itself. Every incoming report needs an origin, a timestamp and a reliability weight: a calibrated flood gauge outranks an unverified social post; a dispatch confirmation outranks a first caller's guess; a belief not reconfirmed within its expected window decays and is flagged for review rather than acted on indefinitely. This is machinery the open position requires, not an objection defeating it. And the sealed annual map keeps a genuine role inside that machinery: as the audited, reproducible baseline from which live streams are compared, not as a substitute for them.
The narrow claim that survives
None of this shows that continuously fed situational awareness is cheap, safe, or currently built at scale in any control room. It shows something smaller and more defensible: on the axis of intake, a sealed annual review, a bounded live feed, and full multi-stream ingestion with provenance are the only three structural positions available. You cannot watch more than every running stream, and you cannot watch it for longer than continuously. Past that point, a system improves by adding sensors, tightening latency, refining decay rates and strengthening provenance — quantitative gains within the third position, not the discovery of a fourth kind of openness. The second law does not prove any of this. It only explains, correctly, why the achievement is never finished: a control room's correspondence with its city is a steady state bought continuously, and the moment the streams are cut, the map relaxes back towards last year's pattern.