The season the assessment expired
In March, a fisheries scientist signs off a total allowable catch for a groundfish stock, built on a stock assessment model last run the previous autumn. The assessment folded in trawl survey indices, catch-at-age data, and a natural mortality estimate that had not moved in three cycles. By July, landings are running 40 per cent ahead of what the assessment implied should be biologically available. Vessels are not cheating. The fish have moved, or bred faster, or died faster, and nobody caught the shift until the September fishery closed six weeks early, log books full of gear conflicts and radio chatter about "fish where there shouldn't be fish."
The proximate cause looks like bad luck: a warm-water incursion nobody forecast, a recruitment pulse nobody sampled in time. The real cause is structural. The quota was fitted to a joint relationship between survey catchability, temperature, and stock biomass that had already changed by the time the number went into the regulation. The assessment was not wrong when it was made. It was wrong by the time it was used, and nothing in the management cycle was built to notice that.
What actually broke
Four streams feed a fisheries management decision, and each has its own update rhythm. Catch reports come in weekly, sometimes with a lag while paper log sheets get keyed. Survey vessels run fixed-station trawls once or twice a year, giving a snapshot that stands in for the whole season. Sea temperature anomalies arrive from buoys and satellite altimetry, often daily, and correlate with recruitment and distribution shifts that outpace the biological sampling built to detect them. Quota filings, the regulatory record of who caught what against what limit, close the loop only after the season the quota governed.
The stock assessment model consumes the survey stream and treats it as current for as long as eighteen months, because that is the interval between survey cruises and the political calendar for setting quotas. Within that interval, temperature anomalies and catch composition can tell you the underlying relationship has moved — fish distribution shifting north, natural mortality rising with a marine heatwave, a year class recruiting stronger than any prior in the time series — but the assessment has no mechanism to listen. It was fitted once, on a fixed data cut, and it stays fixed until the next cruise closes the gap. The quota is a frozen inference wearing the authority of a live one.
This is concept drift, in the specific statistical sense: the relationship between the inputs a model was fitted on and the outcome it predicts has moved, and the fitted rule no longer describes the system generating new data. Detecting it is a formal problem with a real literature, not a matter of scientific vigilance. Walter Shewhart built control charts at Bell Labs in 1924 to separate assignable cause from ordinary noise in manufacturing variance. E. S. Page added the cumulative sum test in 1954, giving sequential change-point detection its lasting form. The idea migrated into machine learning through Schlimmer and Granger's STAGGER system in 1986, Widmer and Kubat's FLORA in 1996, Gama's Drift Detection Method in 2004 — which watches error rate against a Bernoulli confidence bound — and Bifet and Gavaldà's ADWIN in 2007, which compares means across adaptive sub-windows of recent data. Every one of these methods solves the same problem: knowing, from data, when a fitted rule has quietly stopped applying.
Applied to the groundfish case, a Page-Hinkish statistic run on catch-per-unit-effort against temperature anomaly, updated weekly instead of annually, would have flagged the divergence in May — four months before the closure, and in time to revise the in-season quota rather than absorb the overshoot as a fait accompli.
Why the assessment cycle can't do this itself
Detection requires two things the annual assessment cycle does not supply on the timescale that matters: a continuing observation stream, and a way to compare current behaviour against the fitted baseline without waiting for the next full re-fit. The survey gives one clean snapshot a year. Between snapshots, the fishery generates exactly the data — catch composition, effort, bycatch ratios, temperature — that would let a monitor see the drift starting, but the assessment model has no slot for it. It is not that the data doesn't exist. It's that the model's intake stopped at the last cruise, and everything since is invisible to the object making the decision.
The lineage this exposes
A Large Language Model sits at one end of this problem entirely by construction. Its corpus was collected once and frozen at a cutoff; drift after that date is not a limitation it might overcome with more care, it is definitionally outside what the object can observe. There is no stream left running against which to test its beliefs, so none of the drift-detection machinery — CUSUM, ADWIN, Page-Hinkley, prequential error tracking — has anything to act on. Translate that into fisheries terms: it is the stock assessment itself, static until the next cruise, standing in for a fishery that has not stood still.
A Large World Model is closer to the survey vessel mid-cruise. While the scene is present — the trawl is down, the sensors are live, the water column is being sampled — it can detect drift within that episode: a thermocline shift between tows, a net fouling that changes catchability, a school behaving oddly under the hull. That is real detection, and useful. But the monitor ends when the cruise ends. The vessel goes home, the data get processed into an annual index, and the slow drift between cruises — the kind that actually broke the groundfish quota — falls in the gap no episode covers.
A Large Universe Model is defined by the thing this failure was missing: streams that do not stop, and beliefs that carry a timestamp and a source so they can be revised rather than merely replaced. Applied here, it would mean catch reports, buoy temperatures, and survey indices all feeding one continuously updated belief about stock status, each contribution tagged with when it arrived and how much it should be trusted, with an explicit test running underneath asking whether the current catch-effort relationship still resembles the one the last assessment assumed. That is not a better assessment. It is a different kind of object: one for which "the assessment is now two seasons stale" is a measured quantity rather than a discovery made through a closed fishery.
Objections a fisheries scientist will actually raise
Just run the assessment more often. Quarterly re-fits absorb the drift without needing some permanently running belief system.
More frequent re-fitting helps, and many advisory bodies have moved this direction already. But a schedule is open-loop control: it absorbs slow, steady drift and says nothing about whether the interval chosen was ever the right one. A heatwave-driven distribution shift that develops over six weeks will still cost the full quarter before the next re-fit notices it, and nobody will ever know what that cost the stock, because no detector was watching in between. It's also worth noticing that a serious quarterly re-fit pipeline, with retained history and monitored residuals, is already most of the way to continuous intake. It has simply not admitted it.
Detection methods throw false alarms constantly, especially on autocorrelated ecological time series, and the outcome you actually care about — true stock biomass — isn't observable in real time anyway.
This is fair and understates the problem if anything. Verification latency in fisheries is severe: you often only learn the true recruitment of a year class years after the fact, through cohort analysis, by which point any error-based detector has been blind the whole way through. Detectors built on catch and effort alone watch a proxy, not the biological relationship itself, and will sometimes fire on noise. The honest answer is that unreliable detection is still strictly more than no detection, and its unreliability is itself measurable, but only if the stream is kept running long enough to estimate a false alarm rate against. A frozen assessment doesn't have a false alarm rate. It has silence until the closure notice.
The lineage claim is not that fisheries science must ingest every satellite, buoy, and log sheet in existence. It is narrower: that the kind of evidence a model is permitted to observe comes in three exhaustive forms — a frozen corpus, a bounded live scene, or a stream that keeps running with its provenance intact — and only the third can host the detection function drift analysis has always required. Past that point, the argument runs out. What remains is more sensors, better provenance, and shorter latency between a stock moving and someone noticing. That is scale, not a new kind of intake.