Methodology
This page explains, plainly, how a story on this site gets made — and where the process is known to fail.
1. Gathering coverage
Articles are pulled from roughly 28 RSS feeds and a news API on a schedule. Nothing is commissioned or requested from outlets; everything is public reporting they already published.
2. Matching articles into a story
Articles judged to be about the same real-world event are grouped into one story. This matching is automated and imperfect: it can occasionally merge two related-but-distinct events, or fail to match a late-arriving article.
3. Comparing coverage
For each outlet's article in a story, an AI model reads the article and produces: facts it included, a framing summary (headline angle and tone), and language analysis. This is generated text, not a human editor's judgment. The "How each outlet told it" cards on a story page are this model's reading of that one article, nothing more.
4. Extracting claims
The same model extracts specific factual claims from the source articles and rates its own confidence in each one (High / Medium / Low). Confidence is the model's self-rating — it is not computed from source count, cross-checked against other outlets, or independently fact-checked. The "Status" column (Single-source / Multi-source / Contested) is a separate, independently computed field: how many distinct outlets our pipeline found asserting that specific claim.
5. What Tertius does not do
- It does not independently verify facts against primary sources, government records, or on-the-ground reporting.
- It does not silently correct an error in a source article. If an outlet's own reporting contains an internal contradiction — for example, naming a weekday that does not match the date next to it — a claim page flags it visibly rather than fixing it, because the product's value is showing what outlets actually said.
Known limitations, honestly
This is an early-stage, AI-native product. The model can misattribute a claim to the wrong outlet, misread a headline's framing, or repeat a factual error present in a source. If you find one, see Corrections.