How Herald Express Reports: Methodology

How Herald Express Reports: Methodology

Last updated 1 November 2026.

This page describes how a Herald Express article is made, in the order it happens. It describes the system as it runs, including its limits.

1. Finding events, not single stories

Our systems read the public news feeds of many outlets and official sources. Headlines that describe the same real-world event are grouped together. An event is considered for coverage only if at least two independent outlets, or a primary source such as an official body, have reported it. Horoscopes, puzzle hints, live scoreboards and similar low-value items are excluded. We avoid publishing near-duplicates of what we have already covered.

2. Retrieving the reporting

For each event the system retrieves the full text of the reports it is allowed to read. It respects each site’s robots.txt rules, identifies itself, limits how often it contacts a site, and does not bypass paywalls or logins. Reports that largely repeat the same underlying wire copy are counted as one voice, not several. The retrieved material, with its addresses and retrieval times, is the evidence ledger for that article.

3. Drafting from the evidence only

An AI language model drafts the article using only the evidence ledger. It is instructed to attribute claims, to quote only words that appear in the ledger, to avoid unsupported facts and to explain background that the sources provide. Articles must reach a minimum length (1,600 words for standard articles and 3,000 for flagship reports); if the evidence cannot support that depth, the article is not published and is not padded.

4. Independent verification

A second pass checks the draft against the ledger. Separately, deterministic software checks that every direct quotation appears verbatim in the sources, that every number and year is present in the sources, that links point only to sources we retrieved, that the text does not copy long passages from any source, and that no internal markers, lists, tables or diagrams have leaked into the prose. Names that appear nowhere in the evidence are flagged. Any failure either triggers a targeted repair of the affected paragraphs or stops publication.

5. Scoring and publication

Articles are scored on accuracy, depth, writing quality, synthesis, structure, currency and headline. Standard articles need at least 85 out of 100 and flagship reports at least 90, and the score can never override a failed verification check. Articles that pass are published automatically at a measured pace through the day.

6. After publication

The system confirms that the published page loads and carries the right title and canonical address. Corrections requests from readers are handled by people through our corrections process.

Use of AI and human oversight

AI is used for drafting and for part of the verification. Articles are not read line by line by a human editor before publication. People define the rules, audit samples of the output and respond to corrections. See also Editorial Standards and AI transparency.

Known limits

The system can only report what its sources report. It cannot independently witness events. Source errors, ambiguous wording and fast-changing situations can produce mistakes or out-of-date statements. Automated checks reduce these risks; they do not remove them.