Methodology
Each article passes through four research-pipeline stages and a final human editorial review before publication. We publish the intermediate artefacts alongside the article so readers can audit the reasoning.
1. Searcher. Runs 5–9 web queries across different angles (wire coverage, official statements, trade press, local reporting, critical analysis) and fetches the top results.
2. Analyst. Groups results by claim (not by source), tags each claim with a verification level, and writes a source-quality assessment — wire vs. trade vs. blog vs. official — plus an explicit list of framing choices and information gaps.
3. Writer. A local language model composes the article. It must open with the framing comparison and name the missing voices before presenting the factual summary. Every factual claim must carry a source URL.
4. Critic. A second pass checks for unsourced claims, invalid citation URLs, thin analysis, and missing sections. If anything fails, the article goes back for revision or is held from publication.
Every article includes inline citations. No claim stands alone.
URLs are validated against the actual fetched source list.
Quotes are copied verbatim from the fetched extracts.
When a source's framing is noted, we link to the specific article where that word choice appears.
A missing voice must be directly affected by the event, capable of answering a question the coverage cannot answer without them, and specific — "critics" or "the public" do not count.
When we are wrong, we fix the article inline and add a dated correction note at the end. We do not silently edit published pieces. Report errors: editor@sypher-news.com
Sypher News articles are drafted by a local language model following the pipeline above. No third-party AI API sees the research stream. Each article is reviewed by a human editor (Aaron Keating) before publication.
