A cluster of recent commentary points to the same uncomfortable conclusion: several of the most-cited empirical findings of the past two decades do not survive scrutiny as cleanly as the policy debates built upon them assume. The China shock, the rational-choice model of homicide, forecasts of post-SFFA enrollment collapse, and the Huanan market origin story for SARS-CoV-2 are all under revision. None has been formally overturned. All look weaker than their citation counts imply.
Start with trade. According to Joseph Francis, once the corrections proposed by Kirill Borusyak and co-authors in 2022 are fully applied to the canonical China shock analysis covering 1990 to 2007, the estimated damage from Chinese imports to American unemployment, labour force participation, wages, and household incomes disappears. That is not a marginal recalibration. It is the quiet erasure of one of the most politically consequential empirical findings of the 2010s — a finding that shaped tariff policy under two administrations and underwrote a reorientation of the Washington consensus on trade. The literature has not caught up.
Homicides, admissions, and other reversals
A parallel story is unfolding in crime data. Jens Ludwig argues that the post-2019 surge and subsequent volatility in American homicides — overwhelmingly a gun-homicide phenomenon — is poorly explained by the Becker 1968 rational-choice framework still taught as the field's workhorse. Most shootings, on Ludwig's reading, are expressive violence from arguments rather than instrumental violence tied to robbery or drug-territory disputes. He attributes the volatility to a pandemic-era cognitive bandwidth tax that pushed more arguments into automatic cognition, compounded by a large rise and fall in gun carrying tracking handgun sales and police stops around the killing of George Floyd. The broader pattern — homicide trends mirrored by overdoses, car crashes, and suicide, especially among Black Americans — points to a general stress shock, not a crime-policy failure.
Admissions is the cleanest case. The forecasts around the Supreme Court's SFFA v. Harvard ruling — that Black enrollment at selective colleges would fall by 50 to 70 percent, that Harvard's admitted Black share would drop from 14 to 6 percent — have simply not materialised three years in, according to David Kane. The gap between prediction and outcome is wide enough that it deserves the same post-mortem the China shock is now receiving. It has not received one. The asymmetry is itself informative about which errors the profession is willing to audit.
Several of the most-cited findings of the past two decades do not survive scrutiny as cleanly as the policy debates assume.
Two further threads deserve attention. Michael Weissman argues that several lines of evidence suggest SARS-CoV-2 was introduced to humans earlier and elsewhere than the Huanan Seafood Market — a reminder that origin questions remain genuinely open, not closed. And Alexis Akira Toda reports that ChatGPT Pro, Claude, Gemini, and a tool called Refine all generated false positives by endorsing flawed proofs in published mathematical economics papers. That is a specific, operationalisable warning for anyone proposing to deploy large language models as referees or verification tools: the error mode is confident endorsement, not visible failure. The dossier offers no quantified forecasts to anchor a probability spectrum here, but the direction of travel is consistent. Several empirical pillars are softer than their citation record suggests, and the correction cycle is running well behind the policy cycle built on top of them.
Briefings are synthesised by the Ledger Desk from multiple sources cited in the sidebar. They are distinct from Articles, which are written by named contributors and carry a tracked Calibration Index. The Desk does not currently carry a Brier score; this is a deliberate choice for the v0.1 editorial layer and will be revisited.