How we decide what to publish.

Regulator Watch is a rules-based news service. We don’t use a large language model to decide what counts as significant, and we don’t rely on editorial intuition that we can’t describe in writing. Every decision the system makes is reducible to a versioned dictionary of phrases.

The dictionary

The dictionary lives in our codebase at src/lib/filter/dictionary.ts. It has three parts:

• Global include — phrases that mark an item as significant across all regulators (e.g. final notice, licence revoked, cease and desist).
• Global exclude — phrases that mark an item as filler (e.g. speech by, appointment of, investor education week).
• Per-regulator overrides — phrases unique to a single regulator’s house style (e.g. CySEC’s Section 50 public statement, BaFin’s unerlaubte Geschäfte).

Each phrase has a weight. Significance is the sum of include weights minus the sum of exclude weights, and an item passes the threshold if that score reaches our cut-off. The cut-off itself is one number, in one file, that we tune over time.

What we leave out

The exclude bucket is opinionated. We do not publish:

• Speeches and conference appearances — unless they break news.
• Senior staff appointments — unless they are themselves enforcement-relevant.
• Anniversary and milestone posts.
• Investor-education campaigns and financial-literacy weeks.
• Internal operational notices (office hours, system maintenance).

Translation

Non-English items are translated through DeepL into English. The original-language title and excerpt are always preserved and linked from the article page. The original is the authoritative version.

Sources and snapshots

We aggregate from RSS, regulator HTML pages, and (where the regulator’s primary disclosure channel is X) public posts on regulator-owned X accounts. Every item links back to its originating source. We capture and store a snapshot of the source page at the time of ingestion, in case the regulator later amends or removes the page.

Versioning

Every classified item stores the dictionary version that decided it. When we tune the dictionary, we can identify retrospectively which historical items might have been judged differently under the new version. This is unusual for a news service and central to our editorial trust model.

Human review for high-impact items

Severity-4-and-above items — typically warnings, enforcement actions, and licence actions — are held briefly in an approver queue before publishing. Lower-severity items are auto-published. The approver queue is small (typically under ten items at a time) and resolves within minutes.

Corrections

Errors happen. If you spot one, write to corrections@regulator.watch. We commit to reviewing within 24 hours and recording every correction in the audit log.

What is on the dictionary roadmap

We plan to publish the dictionary changelog in human-readable form so industry counterparts can comment. Disclosure of the rules is what separates an editorial product from a black box.

Bank Pulse — separate methodology

The Bank Pulse risk score has its own methodology, including weights, sub-score curves, sources, and algorithm-version policy. See How the Bank Pulse score works.