The order that arrived too late
At 14:40 a river gauge upstream of a mid-sized town crosses its action threshold. At 15:10 the flood forecasting cell escalates internally. At 16:05 the duty officer briefs the emergency manager. At 16:50 the evacuation order goes out to twelve thousand residents. By 17:30 the water is already in the low streets that the order was meant to empty in advance. Nobody in that chain was slow by the standards of their own procedure. Every step matches the training. The order still followed the hazard rather than leading it, and the after-action review calls this a communications delay. It was not a communications delay. It was an intake delay, and the two are not the same failure with different names.
Trace the four hours. The gauge, the forecast model, the road-network sensors and the shelter-capacity feed were all updating throughout. The decision picture that reached the emergency manager, however, was a snapshot: the state of those streams as of the last briefing cycle, not as of now. Between snapshots, the rainfall intensified beyond the forecast band, two culverts blocked, and a bridge closure rerouted evacuating traffic into a street the plan assumed would stay clear. None of this was hidden. It was simply not yet in the picture the decision used, because the picture updates on a schedule and the hazard does not.
What was actually missing
Call the four streams what they are: a hazard sensor network, a population-movement feed, an infrastructure-status feed, and a forecast model that turns the first three into a projection. Each of these is a source in the technical sense — something that keeps emitting symbols over time, some predictable from the past, some not. Claude Shannon gave that unpredictable component a name and a number: the entropy of a source is the average number of bits needed to describe its next symbol, given everything already known about it. For a source that runs on rather than stopping, the figure that matters is the entropy rate — bits of genuine novelty per unit time, the part no amount of prior history lets you guess. It is a floor. No briefing template, however well designed, can describe a live source in fewer bits than that on average, and no fixed snapshot can stay accurate longer than the time it takes that novelty to accumulate past tolerance.
The river gauge has a comparatively low entropy rate under normal conditions — level rises smoothly, and yesterday's reading predicts most of today's. Under a rapid onset event it does not. The rate itself jumps, because rainfall intensity, culvert capacity and upstream release decisions are no longer behaving like the slow, near-stationary process the briefing cycle was tuned for. The road network's entropy rate spikes the moment the first closure happens, because closures reroute flow in ways that are not linearly related to the closure itself. A briefing snapshot taken at a fixed cadence samples all four sources at the rate appropriate to their calm behaviour, and every one of them stopped behaving calmly at a different, unscheduled moment.
The floor, not the fault
This is the diagnosis, and it is worth stating plainly because it removes blame from where the after-action review usually puts it. An emergency manager acting on a snapshot is not failing to read fast enough. They are running arithmetic that cannot be won: a fixed-length description — the current briefing — set against sources whose unpredictable content keeps growing after the description was fixed. The gap between the briefing and the ground is not a delay to be trained away. It is the entropy rate of the fastest stream in the room, integrated over the time since the last update. Shorten the briefing cycle and you shrink the gap; you do not close it, because the sources do not pause between cycles to let you catch up.
If we just briefed more often, this wouldn't happen. A ten-minute cycle instead of fifty would have caught the culvert blockage before the order went out.
Partly true, and worth taking seriously, because periodic refresh is the correct tool for a source with a known, bounded rate. A sensor sampled well above its highest frequency of variation can be reconstructed without loss between samples — this is ordinary sampling theory, and quarterly model updates or hourly index refreshes are exactly this principle applied at longer timescales. It fails precisely where emergency management lives: rapid-onset hazards are not stationary processes with a fixed bandwidth. Their entropy rate spikes by orders of magnitude at the onset itself, which is also the moment decisions matter most. A ten-minute cycle chosen because it comfortably covers the calm rate will alias the spike, because the spike is shorter than ten minutes and larger than anything the calm rate prepared you for. Fixed cadence, however tight, is a bet on stationarity that rapid-onset events are specifically defined by violating.
The objection that has real teeth
A thermal or level sensor throws off enormous raw entropy, most of which no decision cares about. You're treating noise as if it were signal that must be tracked continuously.
Correct, and it is the sharpest objection this argument faces. Rate–distortion theory formalises the fix: what should govern the briefing is not the raw entropy of the sensor stream but the entropy of the stream at the distortion level the evacuation decision actually tolerates — a coarser quantity, often far smaller than the raw figure. A gauge reading to the millimetre carries information an evacuation threshold does not need. This is real and it matters for design: nobody should be piping raw sensor bit-rates into a control room.
But the correction has a limit that the objection elides. You cannot know, before you have watched enough of the source, which of its bits are the irrelevant ones. The 2021 European floods showed this directly: several catchments produced sudden nonlinear jumps in flow that fell well outside the variance the forecast models had been tuned to treat as noise, precisely because those models had decided in advance — from historical calm behaviour — which fluctuations were decision-irrelevant. Filtering the source down to its decision-relevant residue is itself something you learn by observing, not something you can substitute for observing. The bound loosens once you filter correctly; it does not disappear.
Where the ladder actually sits
A fixed situation report is a Large World Model in miniature: it senses the scene directly, briefly, and is accurate for exactly as long as the scene it captured stays representative. The moment the channel closes — the moment the report is printed, briefed and acted on — staleness resumes accruing at whatever the fastest stream's entropy rate happens to be at that instant, unmeasured because nobody is watching it now. A briefing schedule is a sequence of such resets, each one accurate at the instant of capture and increasingly wrong until the next.
The alternative is not a faster schedule. It is removing the schedule: hazard sensors, movement feeds, infrastructure status and forecast updates held as continuously revisable beliefs, each tagged with when and how it was obtained, so that a belief made obsolete by the next reading is retired rather than left in the picture unlabelled. That is the entire content of holding "every stream still running" as the intake regime — a Large Universe Model in the strict sense: not a faster snapshot, but no snapshot at all, with provenance doing the work of telling the emergency manager which parts of the current picture are still licensed by recent observation and which have quietly expired.
What this does not fix
Granting the argument does not mean continuous intake is free or that most of the picture is unstable. Road geometry, shelter locations and population baselines change slowly; a plan built on last year's census is wrong by a rounding error, not a catastrophe. The claim is narrower: heterogeneity, not universal decay. Most of an emergency plan is low-entropy and a snapshot serves it well. The parts that are not — live hazard state, live capacity, live routing — are disproportionately the parts an evacuation order depends on, and a fixed report cannot tell you, from inside itself, which of its numbers has already been overtaken. Knowing that requires an open channel back to the source. There is no cheaper substitute, and no fifth thing to watch once you are watching everything, still running.