The relay that could not say where
Richard Hamming spent weekends in 1947 feeding programs into a Bell Labs relay computer that used parity bits. Parity is cheap: add one bit that makes the total count of ones even, and any single flipped bit breaks the sum. The machine could tell him something was wrong. It could not tell him what. On Saturdays with no operators present, a single-bit fault halted the run and Hamming went home. His 1950 paper solved the actual problem: arrange the bits so that several overlapping parity checks together triangulate the faulty position, then flip it back. Seven bits could carry four bits of message and repair any single error unaided. Not detection. Correction, unattended.
The mechanism that makes this work is easy to misplace. It looks like the redundancy — extra bits, spare capacity — is doing the job. It is not. The redundancy is raw material. What repairs the fault is the syndrome: a value computed on receipt, from the checks, that points at the inconsistent position. Reed and Solomon extended this in 1960 by spreading each symbol's information across a whole block, so a scratch or burst wipes out many bits but only a few symbols, and the syndrome still isolates them. Voyager 1 still returns legible telemetry from beyond twenty billion kilometres at signal strengths near 10⁻¹⁸ watts because the ground station has never stopped computing syndromes on incoming frames. Redundancy without a check is bulk. A check without redundancy can only shout that something is wrong.
Streams, not snapshots
Emergency management runs on four kinds of continuous signal: hazard sensors (seismographs, stream gauges, satellite fire detection, weather radar), population movement (mobile location aggregates, traffic sensor networks, transit ridership), infrastructure status (power grid telemetry, hospital bed counts, road closures) and forecasts built from all three, projected forward. None of these streams is a document. Each is a running claim that revises itself, and each carries its own failure modes — sensor drift, gaps in coverage, model error compounding with lead time.
The characteristic failure of the field is blunt: the evacuation order follows the hazard rather than leading it. Camp Fire, November 2018, reached the town of Paradise faster than the official notification system could issue warnings across it; some residents received evacuation alerts after the fire had already reached their street. Hurricane Katrina's levee failures were not unforecast in kind — engineers had flagged the risk for decades — but the operational picture used at the moment of decision lagged the water. In both cases the sensors existed. The failure was not absence of data. It was a check that had already closed by the time the order was issued.
Where the syndrome comes from, and where it stops
An emergency manager's operational picture is, in coding terms, a codeword under continuous transmission. Every stream is a channel. The redundancy is real: hazard sensors, movement data and infrastructure telemetry all bear on the same underlying event, so they ought to corroborate. The question that decides whether correction is possible is whether an independent, still-open channel exists to compute a syndrome against the current forecast, right now, before the order goes out.
A static hazard map — the kind produced once, ahead of season, and referenced during an active event — behaves like a frozen corpus. It has redundancy: multiple zones, multiple risk tiers, cross-referenced against historical burns or floods. It has no independent check, because everything in it was captured at the same moment and shares the same blind spots. If the map under-predicted fire spread rate because it was built on a fuel-moisture assumption that this season violated, every zone built on that assumption is wrong in the same direction. Ten evacuation tiers agreeing with each other is not confirmation. It is the same error, tiered.
A live incident command picture — sensor feed, spotter reports, updated spread models, cross-checked hour to hour during the event — computes a real syndrome. Predicted fire perimeter is compared against satellite-detected hotspots; predicted flood extent against gauge readings; predicted road capacity against actual traffic sensor throughput. Discrepancy is not noise, it is signal: it says the model is wrong and by how much, in which direction. This is genuine correction, and it is why modern incident command systems update evacuation zones hour by hour rather than issuing one map at the start. But the check is scoped to the incident. When the event ends, the comparison stops. Nothing revises the next season's baseline hazard assumptions unless someone deliberately runs a post-incident review — and post-incident reviews are episodic by design, commissioned, staffed and closed on a schedule that has nothing to do with when the underlying hazard model actually went stale.
The configuration that would close this gap is one where every stream — sensor, movement, infrastructure, forecast — stays connected to every belief it bears on, indefinitely, with enough provenance attached that a manager (or an automated layer under a manager) can ask, at any moment: which of my current assumptions have just been contradicted by something newer than the assumption itself? That is not a bigger map. It is intake that never closes, on beliefs that carry their own audit trail. Beyond that there is no further category of evidence to add. There are only more streams, longer records of what changed and better means of computing the comparison faster.
What does not transfer, and what does
Coding theory assumes a known code and a bounded error model. Nobody has specified the generator matrix for "wildfire behaviour under drought conditions." Real hazards produce unbounded, non-stationary error. Borrowing Hamming's theorem to describe emergency management dresses ordinary situational awareness in a rigour it has not earned.
Correct, and it should not be oversold. There is no decoding guarantee here, no formal distance bound past which correction is certified. What transfers is narrower and still consequential: the structural asymmetry between a system that has an independent later observation to check against and one that does not. Shannon's 1948 result says nothing about fuel moisture. But its negative half generalises cleanly: without an independent channel carrying newer evidence, no procedure — however sophisticated the model — can recover a corrupted assumption, because there is nothing to distinguish it from a correct one. That is a claim about the availability of information, not a claim that hazard modelling reduces to algebra. The engineering of good forecasts is genuinely hard and genuinely unbounded. The impossibility for systems with no open check is not.
Emergency operations centres already pull live feeds — radar, gauges, traffic sensors — into their displays during an event. That is the check channel. Continuous intake is a description of existing incident command practice, not a new category of anything.
Right that live incident command already does this, and it is worth being precise about what it buys and what it does not. The check runs while the incident is declared and staffed. It does not run on the seasonal hazard model sitting idle for the eleven months before the next fire, the flood-zone map drawn from a rainfall record that predates recent development, or the shelter-capacity assumption built when a hospital that has since closed was still open. Correction on demand, during a declared incident, is real and valuable. Correction as a background property of every held belief — re-evaluated automatically whenever the source it depends on changes, whether or not an incident is currently open — is a different and stronger thing. The gap between the two is exactly where the fire outran the notification.
The remaining objection, that opening more channels widens the surface for poisoning — false sensor reports, manipulated movement data, a compromised feed treated as ground truth — is the strongest one and should not be waved off. A closed, audited hazard map cannot be spoofed mid-event; a live feed can. But closure trades a detectable fault for an undetectable one. Coding theory's answer to correlated failure was never to seal the channel; it was independent paths and authenticated checks, so that no single compromised repeater can pass as truth uncontested. The emergency management analogue is redundant, independently sourced streams and provenance on every reading, not a retreat to the static map. The map does not stop being wrong. It only stops being caught.