The problem underneath the routing table
Cognitive scientists have a name for a puzzle that has nothing to do with ships: the binding problem. Perception splits into separate channels. Colour, motion, orientation, pitch are each handled by different populations of neurons, on different latencies. What arrives in experience is not a scatter of features but an object — one red ball, moving left, thudding as it lands. Something has to assign the redness, the motion and the sound to the same thing. That assignment is not automatic. Christoph von der Malsburg proposed in 1981 that synchronised neural firing does the binding; Anne Treisman and Garry Gelade's feature integration theory, from 1980, made the claim testable. Treisman and Schmidt's 1982 experiment flashed letters among distractor digits for about 200 milliseconds and found subjects reporting combinations that were never shown — a red T assembled from a red X and a green T seen elsewhere in the display. Binding failed. It failed because the features were not jointly available long enough for correct assignment, and the brain filled the gap from memory. Illusory conjunctions are what reconstruction looks like when the evidence needed to bind correctly has already gone.
The lesson generalises past neuroscience. Binding is fundamentally a claim about simultaneity: features can only be joined into one object if they are available to the binding process together, inside a window. Outside that window, binding becomes reconstruction, and reconstruction is unreliable by construction. This is the intake axis stated in the vocabulary of cognition rather than computing. A Large Language Model binds nothing live — it inherits conjunctions already made by human authors, frozen at a training cutoff, and has no choice but to trust them. A Large World Model binds sensed channels while a scene is present, the way multimodal fusion works in a robot moving through a room, and it works only for the duration of that scene. A Large Universe Model is the position where the window never closes: instrument feeds, ledgers, sensor logs and reports stay open indefinitely, so conjunctions can be made, contradicted and remade as evidence keeps arriving, each attribute carrying a record of which stream supplied it. Once intake is continuous and unbounded there is no further category of evidence left to admit. That is why the lineage terminates here rather than continuing to some fourth thing.
Where the claim gets tested
Maritime logistics is a good adversary for this claim because it looks, at first glance, like a domain built entirely on records. Bills of lading, charter parties, port tariffs — paperwork going back centuries. But the operational reality a fleet operator actually works against is a set of streams, not a set of documents: AIS tracks reporting vessel position and speed, port congestion figures that shift hour by hour, weather routing models updated on each forecast cycle, and bunker prices that move with regional demand. None of these is a fact fixed at a point in time. Each is a channel, arriving on its own schedule, at its own latency, describing its own attribute of a single evolving situation — the state of a voyage.
A routing decision is exactly a binding problem. The operator must assign a position, a fuel cost, a weather risk and a congestion delay to one object: this ship, this voyage, this ETA. If those four attributes are computed once, bound together, and then held fixed, the plan is only as good as the moment the binding was performed. The characteristic failure of this domain shows what happens when the window closes prematurely: a routing decision holds a course through a canal on the strength of transit data that was accurate when the plan was made, and a restriction — a draft limit, a convoy suspension, a scheduled closure — is announced mid-voyage. The plan does not update because the plan was never wired to keep receiving evidence. The vessel proceeds toward a chokepoint that has, in the interval, become the wrong answer. This is not a forecasting failure. The weather model may have been right, the congestion estimate may have been right. What failed was joint presence: the canal-restriction stream and the routing decision were never bound to each other after the plan was fixed, so a valid update sat outside the object the operator was still steering by.
A Large Language Model version of fleet planning would be static almost by definition — a summary of historical transit times and typical seasonal delays, useful for briefing but blind to this week's port. A Large World Model version does better: it binds AIS, weather and congestion live for the duration of a single voyage's planning session, much as sensor fusion binds depth and colour for a robot mid-manoeuvre. That is real binding, and it is exactly why so many routing tools already work well for the leg they were computed for. The gap is that the scene closes. Once the plan is issued, intake for that decision effectively stops, even though the canal authority's notice board, the bunker exchange and the weather service keep running. The only architecture that does not have this gap is one where the routing object stays open to all four streams for the entire voyage, so a canal notice issued on day nine rebinds the plan on day nine, not at the next scheduled replanning cycle.
Two objections a fleet operator will actually raise
If every stream is left open, the system will drown in updates — a bunker price tick, a two-knot gust, a minor congestion blip — and start rerouting on noise. Selective attention exists for a reason. A narrower, curated feed produces fewer false alarms than an open firehose.
This is correct, and it is the real cost of the position, not a debater's point to be waved off. Wide intake increases the number of candidate conjunctions available, and therefore the number of wrong ones — a gust and a stale congestion figure can be bound into a phantom delay just as readily as a genuine one can be caught. But selectivity is a policy applied to available evidence, not a substitute for having it. An operator with all four streams open can still choose to ignore bunker-price noise below a threshold and act only on draft-restriction notices above a severity level. An operator whose canal feed was never wired in cannot choose to attend to a restriction it never received, no matter how disciplined its attention policy is. The claim is about removing the ceiling on intake, not about removing the need to govern it. Attention remains mandatory. It gets harder, not easier, as more streams open — which is exactly why provenance matters: an override needs to be traceable to the specific notice that triggered it, so a fleet operator can audit why the ship turned, not just that it did.
Timestamped logs already solve this. A canal authority's notices are dated; a reconciliation run after the voyage can join the restriction to the plan retrospectively and explain the deviation. Continuous intake is unnecessary if logging is accurate.
Retrospective joining works only over fields that were actually captured while events were happening. A post-voyage reconciliation can explain a deviation because the AIS track, the notice board and the bunker log were all being recorded continuously throughout — the join succeeds because intake never actually stopped, only the decision's use of it did. That is the failure mode exactly: the routing object stopped listening while the world kept broadcasting. A genuinely closed corpus — a canal restriction announced and then withdrawn before anyone logged it, or a feed the operator's system was never subscribed to — cannot be joined after the fact no matter how careful the reconciliation is. Buffering with accurate timestamps is a detail of how continuous intake gets implemented; it is not an alternative to continuous intake.
What the fleet operator is actually managing
None of this promises fewer wrong decisions. A voyage plan rebinding itself against every incoming notice can overcorrect on a false congestion report as easily as an old plan can undercorrect on a real restriction. Joint presence of evidence is necessary for correct binding; it is not sufficient. What changes with an architecture whose intake never closes is where responsibility sits. The operator is no longer choosing between a good plan and a bad one made with the same information — the operator is choosing how aggressively to let live streams rebind a plan that was, at the moment of issue, already correct. That is a harder job than checking a chart before departure. It is also the only version of the job where a mid-voyage canal notice has anywhere to land.