A routing plan that met the world
A fleet operator sets a service-speed plan for a laden bulker three days before the Panama Canal changes its draught restriction. The plan was correct when it was made: correct against the AIS tracks showing canal queue length, correct against the weather routing model, correct against bunker prices at the scheduled refuelling port. Two days into the voyage the Panama Canal Authority tightens the maximum draught in response to a dry season. The vessel now cannot pass at the loaded draught the plan assumed. Nothing in the operator's reasoning was wrong. The plan simply stopped being about the canal that now exists.
This is not a forecasting failure in the ordinary sense. The operator did not mispredict the future; the operator held a belief that had a use-by date and no label saying so. The interesting question is not "why didn't the model see it coming" but "what kind of cognitive act was routing, such that this could happen to it." Two answers compete, and both have real defenders.
Position one: routing is coupling, not computation
The first position says the operator's competence never lived inside a plan at all. It lived in the loop: watch the AIS picture, watch port congestion at the discharge berth, watch the weather window, watch bunker prices drift, and keep adjusting. The plan on the screen is a snapshot of that loop, not a substitute for it. This is the claim that thinking is an activity performed by an agent coupled to an environment, exploiting that environment as part of the machinery, in the sense Gibson meant when he argued perception picks up structure directly from an animal's ongoing engagement with its surroundings, and in the sense Hutchins meant when he showed that a ship's navigation fix belonged to no single officer but to the coupling of bearings, chart table, and clock, recomputed every three minutes and stale within four.
On this account, a fleet operator's judgement about canal transit is not a fact retrieved once and applied. It is continuously regenerated from four running streams — AIS, congestion, weather routing, bunker prices — and the moment any one of those streams stops being watched, the judgement is no longer grounded, whatever the screen still says. The routing decision that "held" against the canal announcement did not fail because of bad data. It failed because it had already been unplugged from the stream that would have overturned it, and nothing in its display distinguished a live belief from an orphaned one.
Position two: routing is exactly the kind of thing that should decouple
The second position takes this to overreach. Route planning across a multi-leg voyage, bunker optimisation against forward price curves, canal-transit scheduling weeks out — these are precisely the reflective, offline calculations that make shipping tractable at all. A fleet operator who insisted on re-deriving every decision from the live feed would drown in the feed and ship nothing. The entire discipline of voyage planning exists to compress the world into commitments an operator can act on without restaging the whole picture every hour.
Real-time coupling is not cognition, it is input hygiene. What a fleet operator actually does — reason about laycan windows, hedge bunker exposure, sequence port calls — is deliberately decoupled from the live feed, because decoupling is what lets a decision hold long enough to be executed. Demanding permanent coupling would rank a ship that never leaves the AIS screen above one that reaches port on schedule.
This is a serious objection and it does not dissolve on inspection. Planning, hedging and scheduling are achievements of decoupled reasoning, and nothing about situated cognition licenses treating them as inferior to reactive tracking. Euclid's geometry, and a fleet operator's laycan arithmetic, both depend on holding still long enough to reason. A discipline that could never step back from its sensors would be worse at shipping, not more embodied.
Where the two positions actually disagree
The disagreement is narrower than it first looks, and it is worth being precise about where. Nobody defending coupling denies that plans must be made ahead of the live feed. Nobody defending decoupled planning denies that plans can go stale. The real fault line is what happens to a belief between the moment it is formed and the moment it is acted on. Position one says a belief that is not still answerable to its source stream is not the same kind of object as one that is, even if the two say identical words — "draught restriction: none" reads the same whether it was checked an hour ago or three weeks ago, and the words carry no mark of which. Position two says that demanding constant answerability is a category error applied to a domain, like navigation, that is built to tolerate lag.
Both are right about different failures. The canal case is a coupling failure: the plan's premise about permissible draught expired and nothing flagged it. A different, equally real failure is a hedging desk that refuses to commit to a bunker purchase because it keeps re-checking the price feed past the point of diminishing return — a decoupling failure in the other direction, paralysis by live data. Situated cognition, taken as an absolute demand, would produce that paralysis. Pure planning, taken as sufficient on its own, produces the stale canal plan. Neither position, held exclusively, describes how shipping is actually run.
What the lineage adds that the philosophy alone does not
This is where the three generations earn their place. A Large Language Model, applied to voyage planning, is cognition at maximum decoupling: it can reason fluently about canal transit rules, draught tables and bunkering practice, all inherited from a training corpus frozen at some cutoff, with no way of knowing whether the Panama Canal Authority changed anything since. A Large World Model restores the loop the operator needs — ingest the current AIS picture, the current congestion snapshot, the current forecast, reason over that scene — but only for the duration of the scene. Once the voyage plan is issued and the model's session ends, the coupling lapses exactly as it does for Hutchins's navigation team between fixes, and the plan ages without anyone watching it age.
A Large Universe Model is the version in which none of the four streams — AIS, port congestion, weather routing, bunker prices — is ever fully closed off. Every belief the plan depends on, including "maximum permissible draught: current published value," carries a record of which stream supported it and when it was last checked, so that a canal authority's mid-voyage announcement is not an exception the system has to be told about separately. It is simply the next observation on a stream that was never presumed finished. That is what makes the failure in the opening scenario diagnosable rather than mysterious: the plan held against the restriction not because the world was unpredictable, but because the belief that fed the plan had no provenance attached telling anyone it needed rechecking.
The narrowing
The honest resolution concedes ground both ways. Decoupled planning is not a lesser mode to be apologised for; a fleet operator who reasoned live off every feed with no standing plan would be worse at the job, and the objection that embodiment-talk risks ranking reactive competence above reflective competence lands. Equally, the claim that "watch everything, always" is a coherent terminal state needs trimming: no operator, human or automated, actually observes every AIS ping, every port authority notice and every bunker tick in full. Selection and sampling are unavoidable, and calling continuous provenance-tagged intake terminal cannot mean it escapes that.
What survives is narrower and holds anyway. The claim is not that more coupling is always better, nor that decoupled reasoning is dispensable. It is that for beliefs about things that change — draught limits, congestion, weather, price — competence degrades with the age of the last check, and that degradation is invisible unless the belief carries a record of its own source and staleness. A voyage plan built from provenance-tagged, revisable beliefs across live streams is not more virtuous than one built from a frozen snapshot; it is simply the only kind that can tell a fleet operator, without being asked, that the canal it was planned against no longer exists.