How this works

Field Notes publishes findings from new research. Not opinions about research, not predictions, not takes. A finding, what it rests on, and how strong that footing is.

The problem this solves

Ask a language model to write an article with citations and it will invent plausible-looking ones. They have real-sounding titles, real-sounding authors and DOIs that do not resolve. So this system never lets the model choose its own sources.

Four checks, in order

Sources are retrieved, not recalled. Candidate papers come from literature databases with their real metadata attached. A work without a DOI cannot be cited here at all, because it cannot be verified later.
Every DOI is resolved. Each one is looked up against Crossref, and the DOI that comes back must match the one asked for. A fabricated DOI returns nothing. A near-miss that resolves to a different paper is caught by the match check.
Every claim is checked against its sources. This is the part most fact-checking skips. Each claim is sent to a separate pass along with only the abstracts of the papers it cites, and told those abstracts are its entire world. To pass, it must quote the sentence that supports the claim, and that quote is then located in the abstract automatically. A quote that cannot be found is treated as a failure, because a model inventing its own supporting evidence is exactly the failure worth catching.
Retractions are checked. Papers get withdrawn. Citing a retracted study as current evidence fails the article outright.

Evidence is graded, not flattened

A pot trial and a meta-analysis of field data are both “a study”, and treating them alike is how science reporting goes wrong. Every claim carries the strength of the evidence behind it, from meta-analysis down to preprint. Anything weaker than a field trial is labelled where it appears, and nothing resting only on preprints or models publishes without a person reading it first.

What gets published automatically

An article publishes itself only when every reference verified, no reference is retracted, every claim passed its support check, and at least one claim reaches field-trial strength. Anything else stops and waits for a human. The blockers are named, not scored, because a score invites moving the threshold until things pass.

What this does not prove

It does not prove that a paper as a whole supports the claim made from it. An abstract can overstate its own findings. A sentence can be read out of context. That is why every article has a receipts page printing the supporting quote verbatim rather than summarising it: if a check is wrong, you can see that it is wrong.

Corrections are welcome and get made. If a claim here is not supported by the work it cites, that is a bug, and it is treated as one.