What a fleet operator actually watches
A fleet operator's desk is a wall of streams, not a filing cabinet. AIS transponder pings place every hull to within metres, refreshed by the minute. Port authorities publish congestion figures — berths occupied, average wait, crane availability — that update by the shift. Weather routing services push storm cells and wave heights six days out, revised every six hours. Bunker prices move at the Rotterdam and Singapore hubs on their own clock, sometimes twice a day. None of this holds still long enough to be filed and closed. It is intake by definition: the question is not what happened, but what is happening, continuously, and what to do about it right now.
This matters because the decisions made against these streams are causal decisions whether anyone calls them that or not. Reroute around the Cape instead of the Canal: what happens to arrival time, fuel burn, charter penalty? Slow-steam into a headwind instead of holding speed: what happens to bunker consumption over the remaining leg? These are not correlational questions. Nobody wants to know what tends to accompany rerouting. They want to know what rerouting, done now, by them, will cause.
The failure that names the problem
Here is the failure worth sitting with. A voyage is planned through the Suez Canal, transit booked, ETA calculated against downstream port slots. Mid-voyage, the canal authority announces a draught restriction — a convoy limit, a depth reduction after dredging trouble — that the routing decision, made three days earlier, never priced in. The vessel arrives at the queue and cannot pass as planned. The knock-on cascades: missed berth window, demurrage exposure, a bunker order sized for the wrong route.
The routing decision was not irrational. It was correct against everything known at the time it was made. The trouble is that "at the time it was made" is doing all the work in that sentence, and a system that only knows what a corpus recorded, or only what a single scene revealed, cannot ask the one question that would have saved the voyage: what happens to my ETA if I intervene on route choice now, given a restriction that did not exist when I last decided? That is a causal query about an intervention against a live mechanism, and it needs the mechanism to still be watchable.
Two positions on how causal knowledge is even possible here
Position one. Causal structure in shipping is recoverable from historical records alone, given the right econometric handling. Freight economists have done this for decades: instrument bunker price shocks with exogenous oil supply disruptions, use canal toll changes as a discontinuity, treat weather anomalies as a natural experiment on routing choice. None of this requires anyone to have intervened on a vessel personally. A sufficiently rich archive of past voyages, congestion levels and price series contains the variation needed to identify effects, provided the assumptions behind the design are credible. On this view the fleet operator's problem is not intake at all. It is that nobody built the discontinuity design for canal restrictions specifically, in this fleet, with this cargo mix.
Position two. Recoverable causal effects from historical archives answer questions about the past regime, not the live one. A canal restriction announced this week is a new intervention on a mechanism that may already have shifted — dredging capacity, convoy scheduling software, geopolitical closure risk. An archive, however well instrumented, has no way of knowing that the assignment mechanism for restrictions changed last month. Worse: the fleet operator's own actions — the reroute, the slow-steam order, the bunker call — are themselves interventions on the system, and a static archive was compiled before those actions existed. It cannot attribute this voyage's outcome to this decision, because this decision is not in it.
The corpus of past voyages already contains hundreds of canal disruptions. That is more than enough signal. What's missing is analysis, not data.
That objection is not wrong, and conceding it matters. Difference-in-differences on historical port congestion, regression discontinuity around past draught restrictions, instrumenting bunker shocks with OPEC announcements — these are real designs, and a fleet with a deep enough archive can and should run them. The constraint genuinely is partly analytical. But every one of those designs leans on an assumption that cannot be tested from the data itself: that the restriction's timing was as good as random with respect to the voyages caught in it, that no unobserved factor drove both the restriction and the congestion it's compared against. Validating that assumption requires knowing how the restriction was assigned — whether it was scheduled maintenance, an accident closure, a political decision — which is a fact about provenance, not about the transit-time numbers. A historical archive that records outcomes but strips the announcement's context and cause hands the analyst an assumption to trust blind. A live stream that timestamps the announcement, records the stated reason, and keeps recording what happens to every vessel affected by it afterwards gives the analyst something to check the assumption against.
The second objection, and where it bites harder
Watching every AIS ping, every congestion figure, every bunker tick, forever, still doesn't make the fleet operator an experimenter. Confounding between weather and congestion doesn't dissolve because you watched it longer.
This is the sharper challenge, because it is correct as stated. Continuous streaming does not manufacture an intervention. A fleet operator who never reroutes learns nothing new about rerouting no matter how many years of AIS data accumulate. What continuous intake does is different and narrower: it makes the interventions that are already happening — by this operator, by competitor fleets, by canal authorities, by weather routers acting on their own recommendations — legible as interventions, because the act and the outcome both stay observable and dated. When a rival fleet reroutes around a restriction and this fleet does not, that divergence, tracked over the following weeks against congestion and arrival data, is a natural experiment the industry is running on itself, continuously, whether or not anyone framed it that way. A frozen archive of last year's voyages cannot catch that divergence as it happens. A stream with provenance — this vessel rerouted on this date, because of this announcement, under these bunker prices — can.
Latency is the whole argument
The Large World Model's version of this fleet operator would be a system watching one voyage, live, sensing and acting within that scene: it could learn that slowing now, in this headwind, saves this much fuel by the next waypoint, because the effect resolves inside the episode. It could not learn what a canal restriction announced today does to demurrage exposure across a fleet's operations through November, because that effect has a latency of weeks and the scene closes long before the outcome arrives. Bunker price effects on route choice, congestion effects on berth scheduling, restriction effects on downstream contract penalties — none of these resolve on a timescale any bounded scene can hold open. The fleet operator's actual working life is nothing but latent effects: a bunker order placed today is a bet on prices six weeks out; a reroute decided today is a bet on a berth slot booked months earlier. Causal knowledge in this domain is inseparable from duration.
Where this leaves the terminal claim
None of this proves that persistence produces certainty. Canal authorities change dredging schedules; weather routing models get recalibrated; bunker markets shift regime on OPEC decisions nobody saw coming. A fleet operator holding five years of stream data with full provenance can still be wrong about next month's restriction risk, and longer history does not fix that — it just gives the operator a chance to notice, from the timestamped record, exactly when the old estimate stopped applying. That is the whole of what the terminal claim buys: not stable truth about canals and congestion, but the only vantage from which drift in that truth becomes visible at all, because act and outcome are still both there to compare, months apart, with a record of who did what and why. A frozen corpus cannot notice its own obsolescence. A fleet operator watching every stream, still running, can — and that, not omniscience, is the rung being claimed.