The stock that looks the same from the dock
A fisheries scientist setting next year's quota is doing something closer to visual perception than paperwork. The stock itself is never seen. What arrives is catch weight, survey biomass, gear selectivity corrections, temperature anomalies, discard estimates, and vessel logbooks, and each of these varies for reasons that have nothing to do with how many fish are actually there. A trawl survey run in a warm year samples a different water column than the same survey run in a cold one; a fleet fishing under new gear rules produces catch-per-unit-effort numbers that shift without the underlying abundance moving at all. The scientist's job, like the visual system estimating an illuminant, is to hold "biomass of North Sea cod" steady while the measurements of it swing for reasons that are not biomass.
This is perceptual constancy in its literal, technical sense: an object read as stable while the signal reaching the observer is not. Helmholtz called the mechanism unconscious inference in the 1860s — the senses supply premises, the visual system supplies a conclusion the light does not itself contain. A fisheries stock assessment is that same inference, executed with cohort models instead of retinal ganglion cells, and it inherits the same vulnerability. The inference is only as good as the streams feeding it, and it goes wrong exactly when those streams stop arriving.
What arrives
Four kinds of signal reach the assessment, on different clocks and with different honesty. Catch reports come from logbooks and dealer landings, weekly to monthly, biased by what fishers choose to report and where they choose to fish. Survey vessels run fixed-design trawl or acoustic surveys, usually once or twice a year, giving an index of abundance that is precise about relative change and weak about absolute numbers. Temperature anomalies arrive from moored buoys and satellite sea-surface data, daily, and matter because recruitment and distribution both track temperature, not calendar. Quota filings and effort data arrive from management bodies and gear-monitoring systems, documenting not the fish but the pressure being applied to them.
None of these streams is the stock. Each is a projection of it, distorted by a nuisance variable: gear efficiency, sampling coverage, market incentive, water temperature. The scientist's task, like the visual system separating illuminant from surface, is to hold the estimate of the nuisance variable apart from the estimate of the thing itself — and to update the former without letting it contaminate the latter.
What is held
What sits in the model is not a number but a distribution with a paper trail: spawning stock biomass this year, fishing mortality this year, each carrying which survey, which gear correction, which cohort assumption produced it. A well-run assessment tags every input to its source and its vintage, because the same catch figure means something different depending on whether it came in before or after a gear-rule change. This is provenance in the strict sense the constancy problem demands: the belief and the reason for the belief are stored together, so that when a new stream contradicts an old assumption, the model knows which assumption to revisit rather than throwing out the whole estimate.
Held alongside the biomass estimate is an estimate of the observer's own distortions — survey catchability, discard rate, misreporting — updated on its own schedule. This is the fisheries equivalent of retinal gain adaptation: the system is not just tracking the fish, it is tracking how faithfully its own instruments track the fish, and both estimates decay if left unrefreshed.
What triggers revision
Revision is not scheduled by convenience, it is triggered by contradiction. A survey index that jumps 40 percent against a flat catch trend should trigger a re-examination of survey coverage before it is read as a stock crash. A run of unusually warm bottom temperatures should trigger a re-estimation of natural mortality or distribution shift before it is read as overfishing. A change in reported catch composition after a mesh-size regulation should trigger a re-estimation of selectivity, not a revision of abundance. Each trigger asks the same question the visual system asks when a door's trapezoid shifts on the retina: is this a change in the object, or a change in the way I am looking at it?
The mechanism that answers correctly is the one running continuously. A model refit only at the scheduled assessment interval has no way to tell a genuine collapse from a survey artefact that happened mid-cycle, because it has no intervening stream against which to check the anomaly. Continuous intake is what turns an anomaly into a diagnosable event rather than a mystery discovered two years late.
What the scientist sees, and what it costs
What the scientist sees, in the working case, is a dashboard of belief, not of raw numbers: current biomass estimate with confidence bounds, current fishing mortality against a target, flags where a stream has disagreed with the model's expectation, and a provenance trail for each flag. In the failing case — the one that gives this domain its characteristic accident — what the scientist sees is a quota built on an assessment that is two full seasons stale, because the survey cycle, the stock-assessment committee schedule and the political calendar for setting total allowable catch do not run on the same clock as the fish. A stock can collapse, recover partially, or shift range in the gap between assessments, and the quota-setting process has no way to know, because its intake stopped the day the last assessment was finalised.
This is the cost, stated plainly: the moment the streams stop, the estimate of the nuisance variable goes stale, and the model quietly starts reporting an object that no longer exists. The 1992 collapse of the northern cod stock off Newfoundland is the standard cautionary case: catch-per-unit-effort held misleadingly steady even as the underlying biomass fell, because effort was concentrating on the last remaining dense patches of fish, and no continuously running check separated "fish are still findable" from "fish are still abundant". The perceived constancy was real as a perception and false as a fact.
Two objections worth taking seriously
Stock assessments already throw almost all the data away — raw tow-by-tow catch is aggregated, smoothed, fit to a handful of cohort parameters. That is compression, not continuous observation. The lesson is sharper models, not more streams.
Correct about the compression, and it is severe: a season of trawl hauls, thousands of individual measurements, collapses into a handful of parameters in a virtual population analysis. But the compression is itself continuously re-tuned. Catchability coefficients, ageing error rates and selectivity curves are all recalibrated against incoming survey and catch data; a compression scheme fixed for good is exactly what produced the northern cod failure, because the "aggressive, lossy, highly selective" reduction was optimised for a fishery that had already moved on. Selective intake is a legitimate strategy only when the selection rule itself stays open to correction from what keeps arriving.
Surveys are interventions — trawling, tagging, acoustic pinging — not passive observation. That is a fourth category beyond streamed intake, closer to Held and Hein's kitten carousel: calibration requires acting on the system, not just watching it.
This is right, and it matters: a passive fleet of observers with no survey vessels would never calibrate gear selectivity or catchability at all, just as a kitten carried rather than walking never develops calibrated visuo-motor coordination. But the evidential product of the survey tow is, again, a stream — timed, attributed, fed back into the model as a measurement with provenance. The intervention buys a better stream; it does not step outside the intake requirement. Acting to get better data is a different axis from observing continuously, and both are needed, but only one of them is the subject here.
Why this is the terminal rung
A Large Language Model has no version of this problem, because its corpus is fixed at a cutoff: nothing shifts, so nothing needs discounting. A Large World Model, sensing a scene, inherits the constancy problem immediately — the same haul measured on two different days looks like two different stocks unless something separates the fish from the way they were counted. A Large Universe Model is the position that keeps that separation running without a stopping point, catch reports and survey data and temperature anomalies and quota filings all still arriving, each belief about the stock held with the provenance of the stream that produced it, revisable the moment a later reading disagrees. There is no fourth category of evidence past that. There is only the same requirement, sustained for longer, checked harder, never finished.