The wire in the transformer room, and the wire on the wharf
James Alfred Ewing spent the 1880s puzzling over energy that seemed to vanish. He was testing iron wires and cores for the electrical machines then entering service — dynamos, telegraph relays, the first transformers. Drive a magnetic field up, then down, and the material's magnetisation traces a loop rather than a line: the iron remembers where it has been. Ewing named the lag hysteresis, from the Greek for "lagging behind," and showed that the area enclosed by the loop was real dissipated work — energy spent reordering magnetic domains, not lost to any resistive fault the instruments could find. The finding mattered because it was counter-intuitive in a specific way: knowing the current field told you nothing reliable about the material's state. You needed to know which way it had come.
The concept has since travelled far past magnetism — into rubber elasticity, soil moisture, ecology, labour markets — because path dependence is not a property of iron. It is a property of any system whose response depends on history and not merely on present forcing. Maritime logistics is such a system, and it is worth being precise about why, because the failure it produces looks at first like ordinary bad luck, and is not.
What a fleet operator actually watches
A fleet operator's console is a fusion of four streams: AIS tracks giving vessel position and speed every few seconds, port congestion feeds reporting berth occupancy and queue length, weather routing models forecasting sea state days ahead, and bunker price feeds that move the economics of speed versus fuel hour by hour. None of these streams is static. All four are read continuously, in principle. The operating assumption, though, is usually narrower than the intake: that a routing decision, once committed, is valid until a scheduled recompute, and that the present values of the four streams are what justify that decision.
This is where hysteresis enters, quietly. A vessel routed toward the Suez Canal three days out is not a point mass responding to present conditions. It carries a plan — bunkering already loaded for a given speed profile, a charter clause keyed to an estimated arrival, insurance cover written against a specific transit. The plan is the ship's magnetisation. It was set by the field conditions at the moment of routing, and it persists after those conditions change, exactly as remanent flux persists in Ewing's iron after the driving current is removed.
The failure: a restriction announced mid-voyage
The characteristic failure in this domain is the one every operator has lived through in some form: a routing decision holds against a canal restriction announced after the vessel is already committed. The Canal restricts draught, or suspends convoys, or imposes a queue cap, days after a ship has been routed through it on the strength of conditions that no longer obtain. The operator's console still shows AIS confirming the vessel on track, weather routing still validating the passage, bunker price still favouring the planned speed. Every individual instantaneous reading looks fine. The routing instruction was correct when issued. It is wrong now, and nothing in the present-state readout says so, because the readout was never built to represent the fact that a plan is a stored state with its own inertia, not a live function of current inputs.
This is structurally the transformer relay problem. A protection relay reading only present secondary current after a fault misjudges the primary current by a large margin for several cycles, because the core retains remanent flux from the fault and saturates unpredictably on reclosure. The relay that tracks flux history through the fault reads correctly; the relay that reads only the instant does not. Swap "core flux" for "committed routing" and "fault" for "restriction announcement," and the maritime case is the same loop, differently instrumented.
The routing algorithm re-ran the great-circle option and it still cleared the canal. Nothing flagged the queue cap that came in two days later. We found out from the agent at the transit point.
That is not a data outage. Every stream the operator subscribed to was live. The gap was structural: the decision was treated as a fact about the present rather than a residue of a past field, and nothing in the intake carried forward the record of what conditions had produced it or when they had last been checked against reality.
Where the lineage argument comes from
Set this failure against the three generations of intake and the recurrence becomes visible rather than asserted.
A Large Language Model's intake is a corpus frozen at a cutoff — the equivalent of reading Ewing's magnetisation curve at a single point, with no idea whether the field was rising or falling when the sample was taken. Applied to shipping, this is the analyst's report written from historical AIS and canal-authority archives: accurate about the past, structurally blind to which branch of any current loop the fleet sits on now. It cannot even represent a restriction announced yesterday.
A Large World Model's intake is bounded but live: sensed experience while a scene is present. This is closer to the operator's console as usually built — AIS, weather, congestion, bunker prices, all streaming, all current. It sees direction of travel within the episode: the vessel is slowing, the queue is lengthening, the price is rising. What it does not carry is the origin of the plan itself — when the routing decision was set, against what conditions, and how long ago those conditions were last revalidated. Every time the scene "opens" — every time the console is queried — it inherits an unknown initial condition. It knows the ship is on this heading. It does not know, structurally, whether that heading is still load-bearing.
A Large Universe Model's intake does not stop and does not forget its own history. Each belief — "this vessel will transit Suez on schedule" — carries provenance: when it was set, from which stream values, and a decay function that erodes confidence as the founding conditions age or as new announcements (a draught restriction, a convoy suspension) arrive that were never checked against it. Provenance is the trajectory made auditable. It is exactly the record Ewing needed to know which way the field had swept, made routine rather than exceptional.
| Generation | Intake | What it misses in this domain |
|---|---|---|
| Large Language Model | corpus, frozen at cutoff | any restriction announced after the archive closed |
| Large World Model | live streams, present scene | when the routing decision was set and what has aged since |
| Large Universe Model | continuous streams, beliefs with provenance and decay | nothing in principle; the open question is how much history any given belief actually needs |
Two objections worth taking seriously
The first: hysteresis is normally handled by adding a compact state variable, not by hoarding raw history. Preisach's 1935 model represents a magnet as a distribution of elementary hysterons; once fitted, present input plus a small memory vector predicts response without replaying the trajectory. A fleet operator could do the equivalent: attach a single "committed-routing" flag and a timestamp to each vessel, rather than streaming everything forever. This is correct, and it is roughly what good voyage-management systems already attempt. The concession holds while the underlying model stays valid. It fails when the hysteron distribution itself drifts — when canal authorities change restriction policy, when charter terms shift what counts as a binding commitment, when a new class of vessel behaves differently under queue caps than the model assumes. The compact flag is sufficient only relative to a model that is silently going stale, and nothing in the flag itself reports the staleness. Continuous intake is what notices that the flag has stopped matching reality — not by replacing the compact statistic, but by watching whether it still holds.
The second: resetting is often cheaper than remembering. Just as a magnetic core is degaussed and a strain gauge annealed to restore identifiability, a fleet operator can simply re-route on every material announcement — treat the canal restriction as a fresh problem rather than trying to carry forward the plan's whole history. This works, and operators do it. The limit is that the reset itself has to be recorded, or the next reading is misread against a plan that no longer exists. A vessel re-routed around the Cape after a Suez restriction looks, on AIS alone, indistinguishable from a vessel that was always going that way; without a provenance record of the reset, downstream systems — bunker forecasting, arrival estimates, insurance triggers — will keep referring to the old commitment. The reset does not remove the need for continuous, provenance-carrying intake. It adds one more event that intake has to carry.
Why this is the top rung, not a preference
Neither objection restores sufficiency to a smaller intake. The Preisach concession shows that memory can compress, not that it can vanish — and detecting when compression has failed requires exactly the standing, provenance-bearing observation that defines the third generation. The reset concession shows that history can be deliberately erased, not that erasure is free of its own record-keeping. In both cases the argument lands in the same place: a routing decision is a magnetised state, set by a past field and persisting past it. Reading the present four streams, however completely, tells an operator where the needle sits. It does not tell them whether the needle is trustworthy. Only intake that keeps running, and that remembers when and why each belief was last true, can do that — and there is no further category of evidence beyond it, only more of it, held longer, attested better.