Machine-written summaries can be fluent and wrong. Every sentence published
here has to survive four checks before it appears.
1. It must quote its source. The writing model has to attach a
span copied verbatim from the source document to every sentence. If that span is
not genuinely in the source, the sentence is discarded automatically.
2. Its facts must match. Every number, dollar figure, date,
time, vote tally, name and address in the sentence is extracted and checked
against the cited passage. "$1.2 million" and "$1,200,000" are treated as the same
fact; "$1.3 million" is not.
3. A second model tries to refute it. A model from a different
vendor than the one that wrote the sentence sees only the source passage and the
sentence, never the first model's reasoning, and is asked to disprove it. It
defaults to rejecting anything it cannot confirm.
4. High-stakes items get an editorial pass. Anything naming a
private individual, anything from the courts, and the lead item are reviewed
against a checklist covering publication bans, the difference between an accused
and a convicted person, and defamation risk.
The important partWhen a sentence fails
these checks it is removed. When a whole item fails, it is published as a headline
and a link only, marked Headline only. The system is built so that its
worst outcome is telling you less, never telling you something untrue.