Three papers in four do not reproduce from their own package
An automated workflow reran 4,452 published economics replication packages against the papers they belong to and flagged a discrepancy in 3,460 of them.
How we Organise
The price of the same answer halved every quarter for three years
Epoch AI took five benchmarks covering mathematics, hard sciences and games of skill, found the cheapest way to reach each score at each point over three years, and fitted the decline. It comes out at about 47 per cent a quarter, or thirteen times a year. Seventy-five per cent on GPQA Diamond cost thirty cents a question in January 2025, and four hundredths of a penny eighteen months later. Epoch is careful about what that number is. It is a frontier price, and essentially nobody buys on the frontier.
We tell a finance team to come back to a release a month later and put the money it made beside the money it took.
One month after a release, its value is read and set against its cost in the same record.
Over four weeks, at this rate, the frontier price of the same capability moves by close to a fifth. Both rows still exist. The check still passes. What it cannot tell anybody is whether the ratio moved because the team got better or because the input got cheaper, and a finance function reading that record has no way to separate the two. If yours can, we would like to see how it is built. Write to transform@dromologue.ai.
How we Build
The most careful trace any field builds does not give the number back
An open-source workflow took 4,452 published economics replication packages from five journals. For each one it reran the article’s own calculations and compared the result with what the paper reported. It flagged a discrepancy in 3,460 articles or their appendices. These are flags rather than confirmed errors, and a rounding convention or a library version will raise one. The same workflow sped up 496 articles’ calculations more than tenfold and built a fresh extension in 923.
Our position is that a team should be able to produce the same answer twice, and that the record making that possible has to be kept.
Every output is reconstructable: a trace links it to the specification, the versions in play, the context, the tool calls made and the cost incurred.
A replication package is that record at its most deliberate. The author writes it, the journal demands it, and any referee can open it. Three times in four it does not give the paper’s number back. What the study cannot settle is whether an agent’s trace would do better: economics code is not an agent harness, journals ask for different things, and one author worked as a contractor for Anthropic. If you have replayed your own traces and they matched, that is a result worth hearing about at transform@dromologue.ai.
How we Assure
A regulator put 403 million euro on records that were never assembled
Ireland’s Data Protection Commission has fined Google 403 million euro over location data in three of its features, with six months to bring the processing into line. The inquiry opened in February 2020 and covers conduct up to that month, and three of the four findings are about unlawful, unfair or opaque processing. The fourth is different: for one feature, Google could not demonstrate that it had complied. The full decision has not been published, so the regulator’s summary is all anybody has.
We judge a stated commitment by what happens on the day that keeping it starts to cost real money.
A commitment is what somebody is still held to when it costs something.
Six and a half years ran between the conduct and the bill. Everybody who chose those defaults had moved on. That interval is the whole of it. A firm writing its commitments this quarter is writing for a reader who turns up long after the authors have gone, and what that reader may ask for first is the file showing the control ran.