Editorial standards

Every quantitative claim traces to a named producer, and the producer is recorded separately from the service we retrieved it through. A Census Bureau statistic reached through the St. Louis Fed’s API is a Census Bureau statistic; the mirror is named so anyone can reproduce the retrieval, but the authority belongs to the producer.

Every story states the tier of every dataset it uses, on the page, next to the producer’s name.

Before a dataset can be written, it runs a set of deterministic checks. A dataset that fails one is not published, and the failure is an error rather than a warning: the builder stops. The checks that run on every change claim include whether the comparison windows end in the same calendar month, whether the base of the comparison contains anomalous observations large enough to move the finding, whether the claim survives sliding both windows by up to nine months, whether a share’s numerator is genuinely a subset of its denominator, whether every series in a cross-sectional ranking ends on the same date, and whether a relationship survives dropping its most extreme observation.

The checks and their results are published on the story page and inside the downloaded data. That is deliberate. Anyone can say a number was verified; showing which tests were run and what each one returned is the version of that claim a reader can act on.

No number is published that cannot be traced to a source. Missing values are never filled because they look right, and an absent observation is never treated as a zero. An estimate is never presented as a measurement; where an estimate is used, the method and its uncertainty are stated.

No customer data is used, in any form, aggregated or otherwise. The Observatory’s pipeline holds no credentials for any customer system and has no code path that would read one. Where DFX publishes its own measurements, they are measurements taken against public systems, and the method is published with them.

Where a source total is a ceiling rather than a count, it is not published as a count. The SEC full-text index, for example, caps a reported total at 10,000; a query that hits that cap raises an error in our pipeline rather than being published as a floor that looks like a figure.

The Observatory is operated by the General, the autonomous system DFX Intelligence builds. It selects candidate findings, retrieves data, runs the verification, and assembles the pages. That is not a disclaimer, it is the method, and it changes what you should check rather than whether you should trust it.

The relevant safeguard is that the numbers on a page are not written by a language model. Each figure in the prose is computed at build time from the same file the download serves, so a sentence and the dataset under it cannot disagree. Signals are assembled from dataset values rather than paraphrased. Analysis and interpretation are editorial and are written to be arguable; the arithmetic is not.

If something is wrong it is corrected visibly. A correction records the date, what changed and why, and stays attached to the story. Material factual errors are not silently rewritten. Because every dataset is published as a file under version control, the history of a number is recoverable.

To report an error, write to research@dfxintel.com with the page and the figure in question.

Observatory charts and datasets may be reused, including commercially, with attribution to DFX Intelligence and a link to the story they came from. Underlying government data is in the public domain. Cite as: DFX Intelligence Observatory, the title of the dataset, and the date you retrieved it.

Journalists who need a higher-resolution asset, a cut of the data we have not published, or the methodology in more detail can write to research@dfxintel.com.

Candidate findings are generated deterministically from every published dataset and scored on nine dimensions: magnitude, surprise against the rest of the distribution, commercial consequence, audience size, whether a reader can find themselves in it, whether it can be understood graphically, novelty, the authority of its source combined with how much verification it survived, and its usefulness as an answer to a question someone would ask. Candidates below a fixed floor are not published, so a day with three findings worth reading publishes three.

The weights are currently a prior rather than a finding. They are recorded with every published signal so they can be refitted against measured performance rather than adjusted by taste.