What a track file actually is
Air traffic control is the practice of maintaining a continuously updated positional belief about every aircraft in a volume of airspace, and issuing instructions that keep those aircraft apart. That sentence is worth taking slowly, because the word doing the work is "belief," not "data."
A controller's screen does not show aircraft. It shows track files: one record per aircraft, each holding a position, a heading, a rate of climb or descent, a confidence, and a note on where that confidence came from. The track is updated once or twice a second. It is never regarded as settled while the aircraft is airborne. When a radar return is late, the track coasts on its last known vector and the system quietly downgrades its own certainty. When a second source — Mode S, wide-area multilateration, a satellite-relayed broadcast — disagrees with the first, the disagreement itself is logged, not averaged away. The product a control centre sells to an airline is not information. It is separation: a quantified assurance that two objects will not occupy the same point in space at the same time.
That assurance has a price, and the price is spacing. Every unit of uncertainty a system carries about where an aircraft actually is gets converted into a unit of distance it must keep from every other aircraft, and distance is fuel, delay and a limit on how many aircraft can use a given airway in an hour. This is the fact that makes air traffic control worth examining closely before any argument about intelligent systems is allowed near it: it is an institution built entirely around the economics of a belief's freshness, running continuously since the 1930s, with a body count attached to the historical moments when the belief was allowed to go stale.
Flags to satellites
Archie League worked traffic at Lambert Field in St Louis in 1929 with two flags, a chequered one for clearance and a red one for hold. There was no radio; there was no radar; there was a man on a box, watching, and updating his own belief about the field by eye, continuously, because the alternative was aircraft arriving on top of each other. The first purpose-built Airway Traffic Control Centre opened at Newark in 1935. Its technology was a map table, wooden markers pushed across it by hand, and telephoned position reports from pilots who might or might not call in on schedule. It was a belief-maintenance system built from paper and voice, and it worked as long as aircraft were slow enough that a ten-minute-old report was still approximately true.
They stopped being slow enough. Wartime radar entered civil service in the late 1940s; secondary surveillance radar and the transponder, which let an aircraft answer an interrogation with its own identity and altitude, arrived through the 1950s and 1960s. Each generation of equipment solved the same underlying problem in a new register: aircraft speed had outrun the interval between observations, and the gap between an aircraft's true position and the controller's belief about it had become large enough to kill people.
It did, on 30 June 1956, when TWA Flight 2 and United Flight 718 collided over the Grand Canyon in uncontrolled airspace, under see-and-be-seen rules, with nothing between them but filed plans and periodic voice reports — no continuous belief at all. A hundred and twenty-eight people died. The Federal Aviation Act followed in 1958, and continuous radar coverage of the en-route airway structure stopped being an efficiency measure and became a precondition for the airspace being allowed to operate at all. Nobody designed continuous surveillance because it was elegant. It was imposed by the cost of not having it.
The turn
This history describes, without any borrowing from machine learning, a system defined entirely by what it is permitted to observe and when. That happens to be exactly the axis that separates three generations of learning system, and it is worth stating that axis plainly before pushing the comparison further.
A Large Language Model reads a corpus gathered once and frozen at a cutoff; everything after that date is invisible to it by construction. A Large World Model senses a scene while the scene is present — a camera feed, a sensor array, a bounded window of the world — and knows nothing outside that window's edges. A Large Universe Model, as an argued category rather than a built system, holds every stream still running: no cutoff, no edge, a belief per entity that carries its own provenance and decays when its sources go quiet, revised without end because the thing it is tracking has not stopped moving.
Air traffic control is the case where that third position was arrived at first, and arrived at under duress rather than by design philosophy. No controller works from a snapshot of the airspace taken an hour ago. No regulator has ever certified a safety case built on one. The track file — timestamped, multi-sourced, tagged by provenance, downgraded on dropout, never closed while the flight is in the air — is the object the third generation describes, built in the 1950s for reasons that had nothing to do with any theory of intelligence.
What the price tag proves
The strongest form of the claim is that continuous intake is not merely faster batch processing; it is a different category, and the difference shows up as a number. Procedural oceanic control, running on position reports and dead reckoning, required ten minutes of longitudinal spacing between aircraft on the same track. Radar-covered domestic airspace needs three to five nautical miles. When Aireon's receivers, riding on 66 Iridium NEXT satellites, closed the surveillance gap over the North Atlantic in 2019, NAV CANADA and NATS converted the recovered certainty directly into separation: fourteen nautical miles where an hour of flying time had stood before. That conversion — uncertainty traded for spacing, spacing recovered and sold as capacity — is the whole argument in one transaction. The airspace an authority can offer is a direct function of how continuous its belief is, and above full continuous coverage there is nothing left to buy. Gains beyond that come from more receivers, tighter integrity monitoring, longer trusted runs of good data. Not from a fourth kind of evidence.
What this does not license
The weak reading of all this says that more data is better, and that watching everything all the time is self-evidently the goal. Disown that reading explicitly: it is not what the history shows, and it is not what controllers do. A radar screen displaying every return available would be unusable, and controllers are trained to suppress almost everything except the handful of tracks that matter to the decision in front of them. What the institution demonstrates is narrower: a belief that never stops updating, and that carries its own provenance, licenses smaller safety margins than any snapshot ever can. The gain is in what uncertainty stops costing. It is not in volume for its own sake.
Three objections deserve to be taken on directly, because at least one of them cuts real ground from under the claim.
Air traffic control works because the world was engineered to cooperate. Transponders are mandated, routes are filed, aircraft answer when interrogated. Most of the world does not answer when interrogated.
This one is simply true, and it should not be argued away. Mode S, ADS-B Out mandates through 2020 — these were regulatory acts, not technical discoveries. Aviation legislated its own observability. The reply available is narrow: the claim concerns what a system is architecturally permitted to hold as belief, not how easily that belief is filled. Where cooperation cannot be legislated, the same architecture degrades gracefully into lower-confidence, lower-provenance tracks rather than failing outright — which is a real answer, but a smaller one than the aviation case alone suggests.
Returns diminish fast. The North Atlantic flew safely for fifty years on ten-minute HF position reports. Most decisions do not need sub-second truth.
Also true for any single stream taken alone. But the fifty years of safety were bought with the ten-minute buffer, not despite it — procedural separation was expensive precisely because the system did not trust its own belief. The Aireon conversion shows what happens to that expense once the belief improves: it does not vanish, it gets sold as capacity.
A track file is thin — four numbers and a callsign. That a thin belief can run continuously says nothing about whether a rich one can.
Correct, and this narrows the claim properly. Air traffic control demonstrates that continuous, provenance-carrying intake is coherent and economically forced. It does not demonstrate that richer beliefs — meaning, intent, consequence — scale the same way. That richer case has to be argued elsewhere, on its own evidence.
What the concept establishes, then, is the shape of the top rung, not its height. It shows that a belief without a stopping point, tagged by where each piece of it came from, is buildable, survivable, and cheaper than the alternative once running. It does not show that such a belief can be made to hold anything more complicated than where a thing is and where it is going.