The scientist who kept missing the season
The structural problem in fisheries management is older than seismology's telemetry problem but shares its shape exactly. A fisheries scientist assembles a stock assessment: catch-at-age data, survey trawl indices, tagging returns, sometimes acoustic biomass estimates. The assessment takes months to run, longer to review, and by the time a quota is set from it, the fish it describes have moved, spawned, been caught, or died of something the model never saw. The North Sea cod collapse of the late 1990s was not a failure of biology. It was a failure of timing: assessments built on data eighteen months to two years old, delivering quotas timed to a stock that no longer existed in the form assessed. The scientist was right about a fish population that had already changed by the time anyone acted on the finding.
Seismology hit the identical problem a century earlier and solved half of it. John Milne, working in Japan in the 1880s, built the first practical horizontal-pendulum seismograph and, by 1900, ran roughly forty stations worldwide, posting their seismograms back to a base on the Isle of Wight. The record was excellent and it was always late — sometimes by months, limited by the shipping schedules of the era. Nobody used Milne's network to warn anyone. It built a catalogue, and a catalogue was exactly what earth science needed. The Cold War later forced telemetry: the 1959 Berkner panel and Vela Uniform funded the World-Wide Standardized Seismograph Network, 120 matched stations reporting continuously from 1961, because distinguishing a nuclear test from an earthquake required data fast enough and comparable enough to act on before the political window closed. Detection funded the wiring that warning systems later inherited.
Fisheries management has never made that second move. It still runs, structurally, on Milne's model: comprehensive, retrospective, authoritative, and always describing a population that has since moved on.
Three positions, one stock
The lineage from Large Language Model to Large World Model to Large Universe Model is a lineage of intake — what a system is allowed to observe, and when. Seismology already climbed this ladder once, building actual instruments at each rung. Fisheries management has the same three rungs available and mostly occupies the first.
The Large Language Model position is the assessment cycle itself: a frozen corpus of catch reports, survey data and biological samples, closed off at a point, analysed exhaustively, and published as ground truth for a season that has already begun by the time the document exists. It is not bad science. It is structurally incapable of telling a manager that the stock in front of them, right now, is not the stock the document describes.
The Large World Model position is a single research vessel on a single survey leg, streaming water temperature, trawl catch composition and acoustic biomass in real time to the scientists aboard. This is present-tense and grounded — genuinely current information about a genuine scene — but bounded to the patch of ocean the vessel is presently sailing through. A temperature anomaly forming three hundred kilometres north, driving a stock shift the vessel will not reach for another fortnight, is invisible to it.
The Large Universe Model position is the one seismology built and fisheries management has not: every catch report from every landed vessel, every survey leg, every satellite-derived temperature anomaly, every quota filing, streaming continuously into a belief about stock status that is held as revisable, timestamped, and attributed to whichever inputs forced the last revision. Not a faster assessment. A different kind of object — a belief that is always current because it is never finished, with the provenance to say which vessel's catch report, or which buoy's temperature reading, moved the estimate and by how much.
| streams | horizon | characteristic failure | |
|---|---|---|---|
| Large Language Model | closed stock assessment | one season, fixed at publication | quota set on data two seasons stale |
| Large World Model | one vessel's live survey feed | the leg, the patch of ocean | blind to shifts outside the vessel's track |
| Large Universe Model | catch reports, survey feeds, temperature anomalies, quota filings, continuously | none — belief revised as data arrives | none structural; failure becomes institutional |
Nothing beyond the third row is describable in intake terms. There is no fourth category of evidence past "every relevant stream, still running, held as a revisable belief with provenance." What remains, once you reach it, is coverage, cost and trust — which is exactly where the objections below land.
Where the argument gets contested
Real-time catch and temperature data will make the quota-setting body slower and more contested, not faster. Every stakeholder can now dispute the model in real time. Better to fix the number once a year and let people plan around it.
There is real force here, and it maps onto the same problem Japan's earthquake early warning system has lived with since 2007: the earliest estimate is often the worst one. JMA's system has issued alerts for a magnitude 7.8 event that turned out to be 2.3, triggered by near-simultaneous unrelated shocks. A fisheries equivalent is easy to imagine — an early-season temperature spike misread as a stock shift, triggering a precautionary closure that costs a fleet a fortnight's landings for nothing. The honest answer is that this is a criticism of estimator design and revision thresholds, not of continuous intake itself. The fix for a bad early read is more corroborating streams and clearer provenance on each one — did this revision come from three vessels' catch reports or one buoy's anomaly — which is the Large Universe Model position pushed harder, not abandoned. An annual assessment does not have fewer false signals than a continuous one. It has none, because it never looks until the season is already lost.
Even a perfectly current stock estimate does nothing if the quota-setting body still meets twice a year, takes written submissions, and publishes a fixed number that binds for twelve months. The bottleneck is regulatory, not observational.
This is the stronger of the two objections and it should be conceded almost entirely. The 2004 Indian Ocean tsunami is the seismic version of this failure: warning data reached the Pacific Tsunami Warning Center within minutes, and there was no path from that data to a person on a beach in Aceh. Continuous stock monitoring feeding a quota process that only revises annually is the fisheries equivalent — data arriving into an institution with no mechanism to spend it before the window closes. This is exactly right, and it is not an argument against the third rung of intake. It is an argument that intake and actuation are different problems with different ceilings. Actuation can keep improving — in-season quota adjustment triggers, automatic closure thresholds tied to real-time catch rates, the kind of mechanism a handful of fisheries already use for bycatch caps. Intake tops out. Once every relevant stream is flowing continuously into a revisable, provenanced belief about the stock, there is nothing further to observe. What a management body is permitted to do with that belief is a separate, open-ended engineering problem, and fisheries management is mostly stuck on that second problem, not the first.
A third line worth noting briefly, because it will be raised: continuous monitoring is expensive, and the marginal buoy or the marginal onboard sensor buys very little once basic coverage exists. That is true, and it is compatible with the whole argument rather than a challenge to it. Diminishing returns on coverage is what should be expected once the terminal position is approached — the remaining questions become which streams to fund, how much latency in reporting is tolerable, and how much a fleet trusts a number it did not generate itself. None of that reopens a fourth category of evidence. It just confirms that the categories were exhausted before the funding was.
What the recurrence shows
Seismology needed a Cold War procurement problem to fund the wiring that turned Milne's mailed catalogue into a continuous, alarm-capable network. Fisheries management has an equivalent forcing problem sitting in front of it — stocks moving under warming water faster than any annual cycle can track — without yet having the equivalent of Vela Uniform to pay for the wiring. The intake ladder is not hypothetical here. Catch reports, survey feeds and satellite temperature data already exist as separate streams. What is missing is the single revisable belief they could jointly support, and the institutional nerve to act on it before the season it describes has already ended.