Methodology: how our articles are produced

1. Official sources only

The system reads RSS feeds and public APIs of official institutions: central banks (Fed, ECB, Bank of England), statistics agencies (BLS, ONS), regulators (SEC, FCA, ESMA, AMF), the US Treasury, and the contract specifications of CME Group and Eurex. The full list and the collection rules are in the sources policy.

2. Facts, not text

For each topic the system extracts verifiable facts (figures, dates, decisions) with their link, and does not keep the text of the documents. If two sources give different values for the same fact, the topic is held until resolved.

3. AI-assisted drafting

A language model (Anthropic’s Claude API) drafts the article from the fact packet under strict instructions: no figure, date or quote outside the provided facts; no investment advice; no persona; illustrative numbers only inside "hypothetical example" boxes.

4. Automated checks

• factual check: every number and every date in the article must match a fact in the packet or a declared computation that is recomputed;
• a second check by a separate model, claim by claim;
• financial safety: forbidden wording (guarantees, advice, predictions, persona);
• originality: no eight-word sequence taken from a source, no strong similarity with a source or an existing article;
• SEO and editorial review: expected structure, minimum length, no filler.

5. Quality score and publication

Every article receives a score out of 100 (factual 25, originality 20, search intent 15, tool relevance 15, editorial 10, SEO 10, sources 5). It is published only if it reaches the threshold (85 by default) with no blocking failure; otherwise it is revised at most twice, then held for human review. The editor can require manual review before any publication. Volume is capped (a few articles a day): we never publish to fill space.

6. Data and updates

Data panels show official series (ECB, US Treasury, and FRED/EIA when access keys are configured) collected automatically, with their source. One event gives one article, updated when new official releases appear; the update date only changes when the content actually changes. Evergreen articles are reviewed at least every 90 days.

Limits

An automated system can be wrong. The checks greatly reduce the risk of factual error but do not eliminate it. Every article cites its sources so you can verify, and the corrections policy explains how to report an error.