The objection worth taking seriously
An emergency manager running a real operations centre will hear this thesis and push back hard, and the pushback deserves to be stated at full strength before it is answered.
Most of what governs an evacuation call is not moving. The hydraulic behaviour of a river channel, the load rating of a bridge, the burn characteristics of chaparral, the time it takes a hospital to discharge non-critical patients — these are near-stationary facts. They were true last year and will be true next year. A trained manager already carries this knowledge as compressed doctrine: standard operating procedures, pre-modelled inundation zones, tiered response plans keyed to rainfall thresholds. If ninety per cent of the relevant source barely drifts, then continuous intake — every sensor, every feed, held open and re-coded forever — looks like an expensive answer to a marginal problem. Build a good static model of the floodplain once. Update it every few years when the terrain changes. Spend the operating budget on responders, not on infrastructure to watch a river that behaves the same way it did in 1998.
This is not a weak objection. It is, in fact, correct about the average. And the average is not what the cost function of emergency management cares about.
Where the average misleads
Rate-distortion theory measures average distortion over a source, but the losses in emergency management do not distribute like the source. They concentrate on the fraction that moves: the rate of snowmelt into a specific tributary this week, the current occupancy of shelters, which culverts are already blocked, where the fire's spotting distance has jumped because the wind shifted forty minutes ago. The stationary bulk — channel geometry, historical burn patterns, evacuation route capacity — is consulted cheaply, in planning cycles measured in months. The moving tail is consulted under consequence, in windows measured in minutes, and it is precisely there that a frozen description accrues error fastest.
An evacuation order that follows the hazard rather than leading it is the signature failure of this mismatch. The order is a decision compressed from a belief state. If that belief state was coded against yesterday's gauge readings, last week's population distribution, or a road network assumed intact, the order is cheap to issue and wrong at the moment it matters. Rate-distortion theory does not say the stationary majority is irrelevant — it says the distortion measure has to be chosen to match the use, and the use here is weighted almost entirely toward the drifting fraction. Choose that weighting honestly, and the "marginal problem" turns out to hold nearly all the loss.
The retrieval patch, and where it runs out
A second objection follows close behind, and it is the one most operations centres already believe they have solved: retrieval. Attach live feeds to the plan. Pull the current river gauge reading, the latest radar sweep, the shelter occupancy dashboard, at the moment a decision is needed. The static plan stays frozen; the query-time fetch repairs it. This is cheaper than maintaining continuous re-coded belief across every stream, and it is what most emergency operations centres already do with situational awareness boards.
Retrieval genuinely closes a large part of the gap, and any honest account of this domain has to say so. Where it fails is instructive. First, it only retrieves what someone thought to ask for. A manager who queries river stage will not automatically discover that a substation feeding the pumping stations tripped offline twenty minutes ago, because nobody wrote a query for that stream at that moment — the distortion you have not noticed is exactly the distortion retrieval cannot reach. Second, retrieval appends rather than revises. If the static evacuation model says the north route clears in ninety minutes and a live traffic feed says it is currently gridlocked from an unrelated accident, a dashboard that shows both numbers side by side has not resolved anything. It has produced two beliefs where the manager needed one corrected belief, with a record of which number to trust and why. Bolting a feed onto a plan is concatenation. Holding a belief that is continuously re-coded, with the plan's own confidence lowered the moment a contradicting sensor reports, and a note of when that revision happened, is a different operation. That operation is what continuous intake, held with provenance and decay, actually names.
The objection that cuts hardest
There is a third line of attack, and it is the strongest one, because it is not a claim about whether continuous intake is possible but about whether it is affordable, and whether it even helps.
Rate-distortion theory has two axes. Rate is not free. An operations centre that tries to hold every hazard sensor, every mobility trace, every infrastructure status feed open at full fidelity, all the time, is buying bandwidth, storage and compute per bit, indefinitely, against hazards that mostly do not materialise. Worse, live streams are noisy. A flooded sensor housing that reports a false gauge spike, a cell-tower ping mistaken for a mass movement signal, a contractor's status update that lags the actual repair by two days — admitting everything can raise distortion rather than lower it, if every fresh bit is treated as equally trustworthy. Nothing in the mathematics guarantees that the marginal observed bit is worth its storage cost, and an emergency manager drowning in unfiltered feeds is not better positioned than one working from a good static plan. This is an argument about price, not possibility, and it is the one that should be conceded most fully.
The answer is not that observation should be unlimited and unfiltered. Rate-distortion theory itself insists that optimal codes discard aggressively — that is the entire content of the theory, not an exception to it. The claim on the table is about what a system is permitted to observe, not about ingesting every stream at full resolution. A system holding beliefs about a flood basin does not need to store raw radar returns forever; it needs to update its belief about basin state continuously and discard the raw signal once the update is made. Selection has to run inside the system, continuously, deciding which bits earn their keep. Noise is handled the same way a competent watch officer already handles a suspect gauge reading: by weighting it against provenance — which sensor, how recently calibrated, how it has behaved before — not by refusing the feed altogether or by treating every number as gospel until proven otherwise.
What the domain actually needs
Set against these three objections, the narrower claim survives. An evacuation plan built once against historical hydrology is a Large Language Model of the basin: a code purchased at a fixed rate, quoted at a fixed distortion, timestamped at the moment the model was calibrated, never reissued. A common operational picture assembled during an active incident — sensors, radar, road status, all re-coded continuously for the duration of the event — is a Large World Model: distortion is bounded for the scene in front of the manager, and undefined the moment the incident closes and the board goes dark. Neither is what an emergency manager needs across the full cycle of watch, warning, response and recovery, because the hazard does not confine itself to a bounded window and does not wait for a refresh cycle.
What the domain needs is the third arrangement: streams that are never fully closed, beliefs about basin state, population location and infrastructure status that carry a timestamp and a source for every figure, and a decay function that quietly downgrades confidence in anything not recently refreshed. Rate-distortion theory does not license unlimited ingestion — it licenses selection, and selection is cheapest and most honest when it runs against a source that is still speaking. On the intake axis specifically, that is the top rung. Getting evacuation orders to lead the hazard rather than follow it is not a matter of watching more. It is a matter of never letting the watch stop, and never forgetting when each belief was last worth trusting.