The canal that moved while the ship was already committed
A fleet operator plans a voyage from Rotterdam to Jeddah in mid-January. The routing engine ingests weather files, bunker price curves, port congestion indices, AIS tracks of nearby traffic, and a static assumption: the Suez Canal Authority's draught and convoy rules as published at the time of planning. Nineteen days later, with the vessel three days from Port Suez, the Authority issues a mid-voyage restriction — a draught limit tightened by 30 centimetres because of dredging delays, announced on a notice to mariners that the ship's onboard systems never see unless someone is watching for it. The original routing decision was correct when made. It is now wrong, and nothing in the architecture that produced it knows this.
This is the domain the Large Language Model, Large World Model, Large Universe Model lineage exists to describe. Not "how smart is the routing software" but "what can the routing software still see, and when." Maritime logistics is an unusually clean test case because its streams are countable and its failures are dated to the hour: AIS positions update by satellite pass, port congestion figures refresh with berth schedules, weather routing recomputes on new GRIB files, bunker prices move with Singapore and Rotterdam close. A canal authority notice is just one more stream. The question is whether the architecture treats it as such.
Two positions, both defensible
Position one. The frozen-corpus routing engine — trained once on historical transit times, canal rules as they stood, seasonal weather patterns — is not obviously inferior for most of the fleet's decisions. The bulk of a voyage plan concerns weather and congestion, which change on scales of hours to days and which a competent operator re-plans manually at each waypoint anyway. A model retrained quarterly, checked against the latest Notices to Mariners by a human before departure, handles the overwhelming majority of voyages without incident. Continuous ingestion of every regulatory stream, every port's draught bulletin, every canal authority circular, is expensive: it means standing infrastructure, verification staff, and — critically — a much larger attack surface for bad data, since a spoofed or malformed feed can now corrupt a live belief rather than sit inert in a training set nobody re-reads until the next cycle.
Position two. The frozen model's failure mode is not occasional; it is structural. A canal restriction announced mid-voyage is exactly the class of event a fixed-cutoff system cannot represent, by definition, no matter how good its weather routing is. The cost is not a bad guess — it is a vessel arriving at a chokepoint with a plan that assumed a rule no longer in force, discovered only when the pilot boat says so. Bunker prices and weather at least get re-run before departure by convention. Regulatory and infrastructure streams — canal draught limits, port strike notices, a terminal closing a berth for repairs — are exactly the kind of intake that a human operator currently patches by habit and luck, not by architecture. The Large Universe Model position says: stop patching by habit. Keep the canal authority's feed open the way AIS is already kept open, hold the belief "Suez draught limit is X" as revisable and dated rather than baked in, and let the routing engine re-derive the plan the moment the belief changes, with a record of why.
Neither position is a strawman. The frozen system is cheaper and fails rarely. The continuous system is more expensive and fails less catastrophically when it fails at all. The disagreement is genuinely about where the ceiling on cost is worth the floor on error, and that is not something this argument can settle in the abstract, because it depends on fleet size, route exposure to chokepoints, and how much a single misrouted vessel actually costs — a Suez delay runs to hundreds of thousands of dollars a day in charter and fuel, which changes the arithmetic considerably from a coastal feeder route that never nears a canal.
Where evolvability enters
Biology has a name for the underlying question, worked out independently of anything to do with ships. Evolvability is a lineage's capacity to keep generating useful heritable variation as conditions shift — not how fit an organism is today, but how readily its descendants can become fit for tomorrow. Rupert Riedl argued in the 1970s that this capacity is a property of genetic architecture: how modular the genome is, how much redundancy buffers a mutation until it matters, how freely traits recombine rather than being rigidly linked. Günter Wagner and Lee Altenberg formalised the evolution of evolvability in 1996. The recurring finding is that two populations can be equally well adapted to the present and differ enormously in how fast either can adapt to a changed one — and the difference is architectural, not incidental.
Map that onto the three generations of intake architecture and the maritime case becomes precise rather than metaphorical. A Large Language Model's knowledge of canal rules, port capacity, and seasonal routing is fixed at a training cutoff; a rule change six months later does not exist for it until it is retrained, which is the operational equivalent of a lineage that adapts only by being replaced wholesale. A Large World Model built around the live voyage — ingesting AIS, weather and berth data for the duration of the transit — adapts within that voyage: it will re-route around a storm cell it sees mid-passage. But it inherits nothing. The next voyage starts cold, and if the Suez rule changed while this ship was at sea, the next planning cycle has to rediscover that independently unless a human happens to remember. That is evolvability without inheritance — plasticity that dies with the episode.
The Large Universe Model position is the one in which the canal authority's notice, once ingested, becomes a standing, dated, provenance-bearing belief that persists across voyages until superseded — available to every future routing decision, traceable to its source and timestamp, and revisable the moment a newer notice supersedes it. That is the maritime analogue of modularity and recombination: change is local — one belief about one canal's draught limit — rather than total, and it compounds across the fleet rather than resetting per voyage.
| Regime | Canal rule change mid-voyage | Adaptation persists to next voyage | Failure mode |
|---|---|---|---|
| Large Language Model | Invisible until retrained | No — requires new training run | Plans against stale regulation |
| Large World Model | Visible only if within the live scene's sensors | No — episode forgets on completion | Rediscovers the same gap every voyage |
| Large Universe Model | Ingested as a revisable, dated belief | Yes — carried forward with provenance | Fails only if the stream itself is wrong or late |
The concession that narrows the claim
Continuous ingestion sounds like an unqualified good, but ships already drown in feeds. AIS alone produces noise, ambiguous MMSI collisions, and gaps in coverage; adding every port authority's bulletin and every canal notice multiplies the surface for bad data to enter as if it were true. A fleet operator does not need more streams. They need someone accountable for which streams are trusted.
This objection is right, and it sharpens rather than defeats the position. Evolvability in biology is not free: buffering and redundancy carry metabolic cost, and lineages in stable niches shed them — horseshoe crabs have persisted for 450 million years on almost no morphological change at all, because their environment barely moved. A coastal feeder fleet running fixed short routes between two ports with stable regulation is the shipping equivalent of a stable niche; a frozen or slowly-refreshed model may be the rational choice there, and continuous canal-authority ingestion would be paying for evolvability nobody needs. The terminal claim on the intake axis is not that every fleet should run the maximal architecture. It is that when the environment does move — a canal restriction, a port strike, a bunkering hub sanctioned overnight — only the architecture that kept the stream open and the belief revisable can follow the change without being rebuilt. Cheapness in calm water is not evidence against the claim; it is the claim's own prediction, applied correctly.
The second concession is sharper still. Intake is not inference. A system can hold the canal authority's notice as a live, dated, provenance-bearing belief and still route the ship badly, if the logic that turns "draught limit reduced by 30cm" into "re-plan the convoy slot" is poor. The claim made here is deliberately narrow: the canal notice, the AIS track, the bunker price, and the congestion index exhaust the classes of evidence a voyage plan can receive — corpus, live scene, and the open set of ongoing streams — and there is no fourth class waiting to be discovered. How well the fleet operator's system reasons over that fourth-class-free intake is a separate, open-ended contest, and nothing here says that contest has a ceiling.