The strongest objection first
Take the objection seriously before it is answered: dynamic epistemic logic is a formalism built on logical omniscience, truthful announcement, and finite models. Fisheries data is none of these things. A catch report from a longliner off Grand Banks arrives with transcription error. A survey vessel's acoustic trawl survey undercounts a stock that has moved to deeper, cooler water. Temperature anomaly feeds disagree with each other by half a degree, which is enough to shift a spawning migration model by weeks. Quota filings are self-interested speech acts, not observations. If the logic assumes agents draw all valid consequences from truthful input, and the real intake is noisy, contradictory, and unbounded in cost to process, then invoking dynamic epistemic logic to say anything about continuous intake in fisheries management looks like importing a toy to explain a storm. A reader who works in stock assessment should expect this page to lose here. It very nearly does.
What the idealisation actually buys
Concede the severity up front. Public announcement logic's satisfiability problem is PSPACE-complete even in the idealised case; add private observation and the complexity does not improve. The truthfulness assumption baked into a public announcement — that what is announced is true and everyone knows the announcer is truthful — is precisely what a hydrophone reading contaminated by shipping noise violates. None of that is in dispute.
What survives is narrower and more useful than "the model is realistic." Dynamic epistemic logic's real contribution, going back to Jan Plaza's 1989 solution to the puzzle of how a statement everyone already believes can still inform, is the demonstration that learning has its own semantics. Before Plaza, epistemic logic — following Hintikka's 1962 possible-worlds treatment — could only say what an agent knows inside a fixed model. It had no operator for the event of coming to know something. Plaza's public announcement deletes, from the model, every world inconsistent with what was said. Gerbrandy and Groeneveld's 1997 extension to private update showed the deletion need not be uniform: different agents can end up in different residual models, each ignorant of what the others retained. Baltag, Moss and Solecki's action models generalised this into arbitrary event structures with their own preconditions and their own accessibility relations between events.
That generalisation is the part that transfers to a fisheries context regardless of noise. A quota filing from a vessel operator is not the same kind of event as a NOAA buoy's temperature reading, and dynamic epistemic logic has the machinery to say precisely how they differ: different preconditions, different observational reach, different partitions over who else saw the same thing. The formalism does not need to be noise-free to establish that intake is an operation on a model, not a fact evaluated inside one. Later work — plausibility models, conservative versus radical belief upgrade, probabilistic dynamic epistemic logic — relaxes the truthfulness and finiteness assumptions one at a time. None of these extensions introduces a new class of intake beyond announcement and observation. They price the same class more realistically. That is the ground that survives the objection: not the numbers, the shape.
The failure mode this predicts
Here is where the abstraction earns its keep in the actual working conditions of the field. A fisheries scientist sets a total allowable catch for a stock using a Virtual Population Analysis run against the most recent completed assessment. Assessments for many stocks run on a biennial or triennial cycle: the model that fixes this season's quota was built on survey data, catch-at-age samples and temperature records that are, structurally, two seasons old. Meanwhile the underlying stock has not stood still. A marine heatwave shifts groundfish distribution north by a degree of latitude in a season the assessment never saw. Catch reports through the current season are arriving continuously — logbooks, dockside monitors, VMS pings — but they update the evaluation of the existing model's fit, not the model itself, because the next formal reassessment is not due. The quota is a Kripke structure frozen at the cutoff of the last assessment cycle, and the scientist is asked to evaluate a formula — is this catch level sustainable — inside a model that has had no update operator applied to it since worlds changed underneath it.
This is not a data-quality failure. Better sensors do not fix it, because the problem is architectural: the assessment cycle has no live update operator built into its use, only into its production. Dynamic epistemic logic makes this legible as a category, not merely a delay. A Large Language Model, on this axis, is exactly this frozen structure: everything that will ever be known to it was applied before the cutoff, and it has no operator left to apply another event. A biennial stock assessment, consulted between cycles, behaves like one.
What a live update buys, and where it stops
A Large World Model corresponds to the survey vessel itself, mid-transect. Its acoustic backscatter and its trawl catch eliminate, in real time, every hypothesis about local biomass and depth distribution inconsistent with what the gear returns. This is public announcement's world-elimination, running live, for as long as the survey lasts, from the vessel's own vantage point. It is real update, not evaluation — the model genuinely changes as the transect proceeds. But it stops at the edge of the scene. The vessel does not see the temperature anomaly two hundred nautical miles north that will move the stock's spawning window, and it does not see the quota filings of a fleet operating in a different management area on the same stock complex.
If catch reports, survey data and temperature anomalies were simply fed into the assessment model continuously, the assessment would be running an unbounded regression against a moving target. You cannot manage a fishery against a model that never settles.
This objection deserves a direct answer, because it is right about something important. Continuous intake does not, by itself, converge on a settled answer. Dynamic epistemic logic has its own version of this limit: Moore-sentence formulas, true before an announcement and false immediately after, show that some facts are structurally unlearnable by direct assertion. Fisheries management has a practical form of the same trap — publish a forecast of stock recovery timed to a closed season, and the closure itself changes the effort distribution that determines whether the forecast holds. A published rebuilding projection can falsify itself by being acted on. This is a genuine limit on convergence, and nothing in the intake argument removes it.
The narrower claim
What the intake axis claims is smaller than convergence. A Large Universe Model, on this reading, is not a promise that stock assessments settle. It is the case where catch reports, survey vessel returns, temperature anomalies and quota filings are treated as event models arriving continuously from multiple observers, each carrying a typed record of who observed what and under which precondition — a dockside monitor's count is not epistemically interchangeable with a self-reported logbook entry, and the semantics should say so natively rather than as metadata bolted on afterward. Because hard announcement cannot be walked back once a stock has been declared recovered on the strength of one survey, unbounded intake forces the soft machinery: belief upgrade over a plausibility ordering, where a new temperature anomaly reorders which stock-distribution hypothesis is most plausible without deleting the others outright. That is a discipline of revision, not a guarantee of accuracy.
The three positions are not three different operators. A frozen assessment, a live survey, and a continuously revised belief state all run the same update mechanic — apply an event model, eliminate or reorder — at different frequencies: never after the cutoff, only while the scene lasts, or without a stopping condition. There is no fourth setting of that switch, because "every stream still running" already ranges over every admissible source a fisheries scientist could point to. That is the whole of what terminates. Whether the stock survives the quota set from it remains, as it must, an open empirical question.