Where the curse was named
In 1971, three petroleum engineers at Atlantic Richfield — Ed Capen, Robert Clapp and William Campbell — published a paper in the Journal of Petroleum Technology asking why oil companies kept winning offshore leases and losing money on them. The tracts in the Gulf of Mexico drew bids differing by a factor of ten on identical seismic data. The firm that won was not the firm with the best geologists. It was the firm whose geologists had, that week, been most optimistic. Robert Wilson and Paul Milgrom later formalised the mechanics: in a common-value auction, every bidder estimates the same unknown quantity, and the highest bid wins by construction, which means the winner is selected from the upper tail of a noisy distribution, not from its centre. Rational bidders correct by shading their bids downward, in proportion to how many rivals are guessing and how widely those guesses scatter. The work earned Wilson and Milgrom the 2020 Nobel Prize. Richard Thaler carried it into behavioural economics from there, and it now explains overpayment in spectrum sales, free agency, and boardroom acquisitions.
Fisheries management runs the identical auction every season, except the prize is not a lease. It is a quota.
The auction nobody calls an auction
A stock assessment estimates the size of a fish population — cod on Georges Bank, anchoveta off Peru, bluefin somewhere in the mid-Atlantic — from trawl surveys, catch-per-unit-effort records, acoustic surveys and, increasingly, environmental proxies like temperature anomalies that shift a species' range. That estimate becomes a total allowable catch. Fleets, processors and coastal economies then compete for shares of it. The estimate itself is the common value everyone is bidding on, even though no one calls it bidding.
The person responsible for producing that estimate is a fisheries scientist working with a model two seasons out of date by the time it reaches a management council. Survey data take months to process. Assessments take longer to peer review. Councils meet on fixed calendars, not on the rhythm of the ocean. By the time a quota is set, the number it rests on describes a population that may have already moved, spawned poorly, or collapsed under a marine heatwave the assessment never saw.
This is where the curse enters, and it enters twice. First among the scientists: several assessment models may run on the same stock, using different survey subsets or ageing methods, and the one that gets adopted is often the one producing the most defensible-looking, which in practice means the most optimistic, biomass estimate — because a pessimistic assessment invites political resistance and a optimistic one does not. Second among the fleets: vessels competing for the same quota share act on whichever recent catch reports and temperature signals make the stock look most available, and the operators who commit capital hardest are the ones whose private read of the water was most favourable. Both are selections on the maximum of a noisy sample. Both are cursed by the same arithmetic Capen, Clapp and Campbell wrote down in 1971.
The distinguishing failure mode of fisheries management is not that assessments are wrong. It is that they are wrong in a directionally predictable way, and the direction is optimism, because optimism is what survives the selection process that turns a scattered set of estimates into one adopted number.
Why a frozen assessment cannot shade its own bid
A stock assessment behaves like a frozen corpus. It is built from data up to a cutoff, reviewed, published, and then acted upon for one to three years until the next cycle. Inside that window, the assessment carries an unknown and growing staleness term. A council setting quota in year two of a three-year assessment cycle is, structurally, bidding on last year's estimate of this year's fish. It has no principled way to know how much the population has drifted since the survey vessels went home, because the thing it would need to know — how much has changed since the observation — is precisely what a frozen assessment cannot contain. It was current once. It does not know its own age in a way that lets it correct for it.
Bid shading does not require live data. A well-calibrated assessment model with an uncertainty band can shade the quota correctly without observing anything new — that is what confidence intervals are for.
This objection has real force, and stock assessments do carry confidence intervals. But those intervals are typically calibrated to the average staleness across the whole assessment cycle, not to the specific gap between the last survey and the current quota-setting meeting. A population undergoing a recruitment collapse in year two of a three-year cycle needs a much wider band than a stable population in year one, and a fixed-width interval cannot tell the difference, because knowing which situation you're in requires exactly the kind of recent observation the assessment was frozen before receiving. Correct shading is conditional on elapsed time since observation, and elapsed time since observation is only knowable by watching the clock run — which is intake, not calibration.
Where survey vessels alone fall short
Real-time acoustic surveys and satellite-tracked temperature anomalies improve on this, and some fisheries now run near-continuous monitoring for exactly this reason. A survey vessel tightens the estimate for the patch of ocean it is currently sampling. This is the Large World Model move: bounded, current, and sharply informative inside the scene the sensors cover.
It says nothing about the water forty miles north, where the same stock may already be shifting under a warm-water anomaly the vessel hasn't reached. Bluefin tuna and mackerel have both shown multi-hundred-kilometre range shifts within single seasons as thermal fronts move; a survey plan designed around last decade's known range samples the wrong water entirely, confidently, and reports high confidence about the wrong thing. Tightening error inside a scene does not extend to the boundary of the scene, and quota decisions are made about the whole stock, not the sampled patch.
The shape of the correction
What the curse actually demands is not more data and not fresher data alone. It demands knowing, for every input feeding the assessment, its age, its source and how far it has already been revised — enough structure to compute a conditional variance rather than assume a fixed one.
| Input | What it typically has | What shading requires |
|---|---|---|
| Trawl survey | A sample date | A decay curve since that date, by species mobility |
| Catch reports | A filing timestamp | A reliability weight by fleet and by gear type |
| Temperature anomaly | A satellite pass | A causal link strength to the specific stock's range |
| Prior assessment | A publication date | An explicit staleness term carried into the next cycle |
A system that maintained this — catch reports, survey passes, temperature anomalies and quota filings, each tagged with age and source and revised as later data arrived, rather than replaced wholesale every assessment cycle — would be able to compute how much to shade this season's quota below the point estimate the raw data suggests. That is not a better stock assessment model. It is a different intake architecture: continuous, provenance-tagged, revisable belief rather than periodic frozen snapshot.
Continuous intake just adds more noisy streams. More sensors, more spurious anomalies, more chances that one glitchy buoy reading gets treated as signal and the curse gets worse, not better.
This is a genuine risk and it has happened: a single malfunctioning temperature buoy has driven range-shift assumptions into an assessment before being caught. The fix is not less intake. It is that each stream's historical reliability is tracked and used to downweight it automatically, so one anomalous buoy is discounted rather than allowed to set next season's quota. A frozen assessment cannot do this at all, because it has no revision history to weight against — every number in a published stock assessment carries equal typographic authority, whether it came from twenty years of consistent survey coverage or one contested tow.
The lineage this recurrence points to
A Large Language Model, asked about a fish stock, would answer from a corpus with an unknown and growing gap since its cutoff — it cannot shade because it cannot see its own staleness. A Large World Model, built from live survey and satellite feeds, tightens the estimate for the patch of ocean currently observed and leaves everything outside that patch exactly as uncertain as before, often without saying so. A Large Universe Model, defined by carrying age, source and revision history across every running stream — catch reports, survey passes, temperature anomalies, quota filings — is the only architecture that can compute the conditional variance the correction actually requires.
This is not a claim that such a system exists in any deployed fishery today. It is a claim about what the arithmetic of the winner's curse permits as a correction, and about where, on the axis of intake, that correction becomes computable rather than merely hoped for.