← Back to the desk

Methodology

Sources

SignalDesk uses only official RSS, API or open-data endpoints whose reuse terms are documented. Each source has a rights status. Verified sources can be activated once a live endpoint check passes. Needs Review (the default for anything unknown) and Blocked sources are never fetched. Feed URLs are taken from the publisher's own directory, never guessed. The first source is the Reserve Bank of India's RSS directory.

What we store

Headline, link, publish date and an excerpt of up to 500 characters, attributed to the publisher. Full articles are not stored, paywalls are not bypassed and restricted pages are not scraped. Readers follow the link to the original.

Deduplication & topics

Links are normalised (https, no www, tracking parameters removed). A record is a duplicate if its normalised link or its fingerprint (normalised title + publish day) already exists. Topics are assigned by keyword rules across nine topics, falling back to the source's default.

Claim labels

  • VerifiedStated directly by an official primary-source record.
  • ReportedStated in a record but not independently confirmed.
  • EstimateA figure, range or direction derived from records.
  • AI AnalysisThe model's interpretation. It still has to cite the records it reasoned from.

Every claim must cite at least one stored record. The server removes any claim whose citations don't match records supplied to the model, and shows how many were removed. Each analysis has an uncertainty note.

Projections

A projection needs at least three records from the last 30 days in the chosen topic. With fewer, SignalDesk refuses before any AI call. The model may also decline when the records don't support a view.

Publication

Briefings and weekly editions are private. A public piece moves from draft to in review to approved to published. Only a signed-in human editor or admin can approve or publish, and the database enforces this. Automated jobs cannot publish. Editing an approved piece sends it back to review.