Methodology

Clustering

We group reporting that addresses the same underlying event into a story cluster. A neutral headline and summary describe what is common across accounts. Clustering is reviewable: editors can merge related clusters or split one that has drifted apart.

• A cluster requires at least two independently sourced accounts.
• Editors confirm the accounts address the same event before clustering.
• Related clusters are linked, never silently merged.

Summaries

Neutral summaries are AI-assisted drafts, then reviewed by an editor before approval. A summary describes the shared event and the parameters of disagreement. It does not resolve disputes the sources do not resolve.

• Approved summaries are labeled as such; pending summaries are labeled under review.
• Editors can edit any summary. Every edit is recorded in the audit log.
• Summaries never assign motive or rank outlets.

Confirmed facts and disputed claims

Confirmed facts are statements corroborated by multiple indexed sources, each linked at the claim level. Disputed or developing claims describe the disagreement and what is incomplete — they are not declared true or false.

Coverage comparison

For each cluster we record observable differences across publishers: headline, lead emphasis, quoted voices, sourcing, omissions, emotional language, and framing. We describe what differs. We do not produce a bias score, because a score implies a single correct framing we do not assert.

Differences are not equal to intent. An omission may reflect space, audience, or editorial priority — not necessarily a deliberate choice.

Lean assessment

Each source carries an editorial lean — Left, Center-Left, Center, Center-Right, Right, or Mixed — shown on its profile and as a small dot beside its name in story coverage. This is a label, not a computed score: it is assigned by an editor and backed by a cited basis (a published methodology such as AllSides or Ad Fontes, refined with our own observed coverage) and a short rationale. Unassigned means we have not yet assessed the source.

• Lean is a conclusion backed by visible evidence, not an opaque algorithmic rating.
• Every assignment records its basis and rationale, visible on the source profile.
• Changes to a source's lean are logged and appear in Corrections.

Labels

Every piece of content is labeled: Reporting, Analysis, Opinion, Automated Analysis, or Sponsored. Sponsorship never influences clustering or comparison.

Tabloid-overload detection

When outlets known for sensationalized, emotionally charged coverage publish disproportionately about a story, we flag the cluster with a tabloid-overload banner. This is a count-based signal, not an editorial judgment of the story itself — it tells readers that the coverage mix is weighted toward tabloid sources and warrants extra scrutiny.

• Core tabloids (e.g. New York Post, Daily Mail, news.com.au, Times of India) count at full weight; a single outlet publishing 2+ articles is enough to flag.
• Borderline outlets that mix sensationalism with legitimate commentary (e.g. Daily Wire, HuffPost, Daily Beast) count at half weight, so a cluster needs roughly twice as many to trigger the same signal.
• The banner names the specific outlets involved and their share of coverage, not a score.