What arrives, and how fast
An incident commander running a fire in steep, fuel-loaded terrain is a receiver at the end of several channels, each with its own symbol rate. Fuel-moisture sensors on masts and in soil report every ten to sixty minutes: dead fine fuel moisture, live fuel moisture, an index derived from both. Satellite thermal passes arrive on an orbital clock, not a tactical one — a polar-orbiting sensor might revisit a given ridge every twelve hours, sooner if a second satellite's swath overlaps. Wind models — mesoscale forecasts nested down to hundred-metre grids — update on the hour, sometimes every fifteen minutes near a fire perimeter when a modelling centre prioritises the event. Crew positions come off GPS trackers on trucks and, where issued, on individuals, reporting every few seconds when signal holds and not at all in a canyon shadow.
None of these streams is continuous in the strict sense. Each has gaps, and each gap is where Claude Shannon's 1948 result becomes an operational fact rather than an abstraction. Capacity is a rate. A sensor that reports every hour has, between reports, an instantaneous rate of zero about the fuel moisture on that slope right now. The question an incident commander actually works is not "what do I know" but "at what rate is my knowledge updating, on which channel, with what lag" — because the fire does not wait for the next scheduled report.
What is held
The command post holds a running situational picture: a fire perimeter polygon, a fuel-moisture surface interpolated between sensor points, a forecast wind field for the next six, twelve, twenty-four hours, and crew markers on a map. This is not a fresh capture of the present. It is a stitched object built from streams closing at different times. The perimeter might be four hours old, drawn from the last useful thermal pass. The fuel-moisture surface might be interpolating readings that are, on average, thirty minutes stale at any given mast and much staler in the gaps between masts, which is most of the terrain. The wind field is a forecast, meaning it was never a measurement of the present at all — it is a projection whose own confidence decays with lead time.
Each element carries, or should carry, provenance: when was this measured, by what instrument, with what stated uncertainty. A fuel-moisture reading from a mast three kilometres from the flank is not the same evidentiary weight as one on the flank itself, and a commander who treats them identically has collapsed provenance into a single undifferentiated map, which is precisely the failure mode that makes the picture look more current than it is.
What triggers revision
Revision happens when a new symbol contradicts the standing belief, not merely when it updates it. A wind model bulletin at 14:00 forecasting a shift from southwest to northwest at 16:30 is a revision trigger of the first order: it invalidates the flank assignment made an hour earlier, because a flank drawn against a southwest push becomes a head against a northwest one. A thermal pass showing a spot fire eight hundred metres beyond the mapped perimeter is a second-order trigger: it does not change the forecast, it changes the ground truth the forecast was meant to apply to.
The characteristic failure of this domain is that the trigger arrives late relative to the event it should have preceded. An ignition is detected after the wind has already shifted, rather than before. This is not a sensor failure in the narrow sense — the satellite passed, the model updated, the data are in the log — it is a duty-cycle failure. The channel that should have warned of the shift runs on an update cycle measured in tens of minutes to hours; fire behaviour on a wind shift changes in single minutes. Rate of intake about wind state is bounded below the rate at which wind state actually changes at the surface, on the specific slope, in the specific chimney or saddle that channels flow. The gap between those two rates is where the loss happens, and no amount of resolution on the eventual thermal image closes it, because the image arrives after the run.
"We had good data. The 16:30 update showed the shift clearly. The fire crossed the road at 16:40."
This is the objection stated in its own voice, and it deserves to be taken seriously: the data existed, the model was not wrong, the commander was not negligent. The failure is structural — a ten-minute-resolution wind product cannot warn of a shift that reorganises fire behaviour inside that same window. Improving the model's accuracy does not fix a rate mismatch. Only raising the rate does, and raising the rate has a cost measured in instrumentation, bandwidth and compute that a hundred-thousand-hectare incident may or may not have been allocated.
What the operator sees
What reaches the commander's screen is a further compression of all this: a common operating picture, updated on whatever cadence the incident management system supports, typically minutes for position data and hours for perimeter and fuel layers. The interface cannot show provenance in full without becoming unreadable, so most systems collapse it to a colour or a timestamp — green polygon, last updated 13:42. The commander must reconstruct, from experience, what that timestamp means for trust: a 13:42 perimeter on a slow-moving fire in stable air is nearly as good as live; the same timestamp on a fire that has just crested a ridge into unstable air with a frontal passage forecast is close to useless, and known by the experienced commander to be close to useless, which is itself a form of tacit provenance-tracking the formal system does not encode.
This is where duty cycle becomes a lived quantity rather than an engineering term. A commander who has been on the line for eighteen hours is not merely fatigued in the ordinary sense; they are integrating streams at rates that do not match each other and trying to hold a coherent, revisable belief about a system — the fire, the wind, the crews — that keeps emitting whether or not anyone is watching. Closing one's eyes does not close the fire's channel. It only closes the commander's.
What it costs
Raising duty cycle is not free, and this is the objection that carries the most weight against any argument that "more streams, more often" is simply better. A network of fuel-moisture sensors dense enough to resolve slope-scale variation, reporting every five minutes instead of every sixty, multiplies bandwidth, battery replacement cycles, and the analyst-hours needed to keep the feed usable rather than merely present. A thermal-imaging pass frequent enough to catch spot fires within minutes rather than hours requires tasking a satellite constellation or flying a dedicated aircraft, and aircraft compete with retardant drops for the same airspace and the same budget. Crew trackers reporting every second drain batteries faster in exactly the conditions — heat, exertion, distance from a charging point — where a dead tracker is most costly.
None of this is an argument for maximal ingestion. It is an argument that intake is always an allocation problem, and Shannon's framing makes the allocation explicit rather than implicit: capacity spent on one stream is capacity not spent on another, and a channel run at its limit degrades under the noise any real deployment introduces — canyon shadow, smoke attenuation on thermal bands, GPS multipath near steep terrain. A commander's real question is not "can I get everything" but "which stream, raised in rate, would have prevented the last failure, and what do I stop watching to pay for it."
What retrieval and forecasting do not fix
A live weather model queried on demand is sometimes offered as proof that the closed-channel problem is already solved — pull the latest run whenever needed. But a forecast queried at 13:00 answers the question posed at 13:00, using priors about atmospheric state formed before that query; it cannot report a shift that begins forming at 13:05 unless the query is reissued, and reissue frequency is a duty cycle choice made by the commander, not a property the model supplies for free. This is the retrieval objection in its wildfire form: pull raises rate only as fast as the puller pulls. The terminal position on this axis is not better pulling. It is a channel that stays open regardless of whether anyone remembered to ask, with every stream's provenance intact so that the 16:30 report can correct the 16:00 belief before the fire, not the debrief, does the correcting.