Ending A · The audit
If the gatekeepers make verification mandatory and rewarded, 2029 to 2031
This is one of two endings. It begins from the same premises as Ending B · The flood, and the two paragraphs below are identical in both. It diverges only on the gatekeepers’ choice. Here, they choose to reward verification.
By the end of 2028 both endings start from the same position. Generating plausible economics is cheap and getting cheaper. Checking it is not, and the corpus shows the field’s first response was to build filters. Verification is the second largest thing the collection’s tools are built for, labelled by hand, and the most internally varied of them. Agents rewarded the projects that were versioned, tested, and documented, so software engineering discipline arrived through the back door. The embargoed research frontier that held coverage down does not open by itself. Mirrors appear on authors’ own sites as embargoes lapse, but the index only reflects them when someone runs the fetch again. Talent equalized as the cost of doing the work fell. Access stratified as infrastructure decided who could do it at all. The profession is at once watching cheap intelligence and being reshaped by it.
One decision has not resolved. Whether the field’s gatekeepers, meaning journals, funders and data editors, make machine verification mandatory and rewarded is not a fact about the models. It is an institutional choice, and it is the choice that sorts the next three years. The models keep improving in both endings. The difference is entirely whether verified judgment, the scarce factor, gets priced and rewarded before the flood of plausible machine text outruns the humans left to supervise it. What follows changes only that one variable.
In this ending the gatekeepers say yes, and they say it early enough to matter.
2029, the gate closes, and it holds
Replication checked by an agent becomes a condition of publication rather than a courtesy. The move starts where the machinery already existed, in the data editor function that had spent a decade making code and data availability a requirement across the discipline, and it spreads because the tooling to enforce it is now cheap. A submission arrives with its pipeline. The pipeline reruns from raw data to every table. A reproduction that fails is a desk reject. The reporting standard that got a name in 2026 is now the thing a methods section cites, the way a robustness section is cited, and the referee factory the field built for itself becomes the referee infrastructure the journals lean on.
The cost of producing a result keeps falling, so more people can get in. A researcher without a large lab can run an agenda that would have needed one. Talent equalizes. This is the good news, and it is real.
2030, the unit of contribution moves
By 2030 the published literature is measurably more reliable than it was in 2020, for a plain reason. Every empirical claim that reaches print has survived an automated reproduction the 2020 literature never faced. What is hard to fake gains value accordingly. A paper is easy to fabricate. A maintained dataset, a tested codebase, and a living document that reruns end to end are not, so the unit of contribution drifts toward them. The thing people cite becomes an object, a benchmark, a standard, a package, rather than a single paper.
A market grows in the space the field had flagged as empty. Nobody had been modelling the market for verification itself, and now audit, certification, provenance, and liability for machine outputs become questions with funding attached. The assurance industry that regulation is calling into being hires economists to price model error and design certification markets. The escape route the corpus kept pointing at, which is to keep an audited ground truth check somewhere in the chain, becomes the normal way to build things rather than a habit of the careful few.
Verified judgment got priced. The scarce factor found a market, and the market paid for the checking the flood would otherwise have skipped.
2031, audit first, and its cost
The audit layer is plumbing by 2031, as invisible as preregistration became. A cohort taught to check first, graded on finding planted errors, examined out loud on work that cannot be handed to a machine, enters the field able to supervise machines it never had to do the grunt work to understand, because the curriculum rebuilt the judgment the apprenticeship used to supply.
The honest cost sits underneath the good news. Talent equalized, but access stratified. The things that make an audit cheap, meaning compute, restricted data enclaves and maintained toolchains, are unevenly held, so the gap that widened is between institutions rather than between people. The field is more reliable and less evenly staffed than it was, and both follow from the same choice.
What would have to be true by 2030 for this to be the ending unfolding
The first marker is that at least one top five journal or the AEA Data Editor runs mandatory reproduction by agent on submissions, with a failed reproduction treated as grounds for rejection, as policy rather than as a pilot.
The second is that a maintained, cited reporting standard for machine assisted measurement is in routine use, appearing in methods sections and backed by at least one maintained validation suite rather than a scatter of personal repositories.
The third is that hiring and teaching reward the audit skill, so research assistant and predoc postings ask for verification and replication competence, and audit style assessment appears in graduate syllabi.
If instead refereeing capacity broke before the gate could close, you are in Ending B · The flood.