Methodology
The lean labels and perspective analysis on this site are generated by AI, applied to individual articles, and updated with every feed refresh. Here is how the system works and where it falls short.
How article lean is scored
When articles are ingested from RSS feeds, each article is analyzed by an AI language model. The model reads the headline and article content and assigns a lean score on a scale from −1 to +1:
For national/Bangla coverage. Negative scores (Left) mark framing that amplifies the government's own voice; positive scores (Right) mark framing that sides with the opposition — Jamaat and pro–Awami League forces — or holds the government to account. Most articles score 0 and stay Independent.
The AI model considers signals including: whose perspective is foregrounded, whose statements are quoted first or most prominently, how events are characterized (protests vs. unrest; crackdown vs. security operation), and whether criticism of a party or institution is present or absent.
The model also returns a confidence level for each score. Articles scored at lower confidence are shown with the broader "Center" classification rather than a strong lean call.
Bangladesh versus international axis
Bangladesh's politics don't map cleanly onto the Western left/right spectrum, so for domestic coverage we measure something more useful: a story's stance toward the current government. Since the February 2026 election the BNP governs, the Awami League is banned, and Jamaat-e-Islami leads the opposition. A story leans Left when it amplifies and promotes the government's own voice — a minister claiming credit, reassuring the public, rebutting rumours (the BNP governs with a two-thirds mandate). It leans Right when it centers the party-political opposition — Jamaat and pro–Awami League forces — or holds the government to account. Everything else — plain reporting, the courts, routine administration, crime, business, culture, and all foreign news — is Independent, the default. The axis judges stance toward power, and a neutral quote from any politician stays Independent.
We use two separate axes accordingly:
International outlets (BBC, Reuters, Al Jazeera, The Guardian, DW, NYT, Fox News, France 24) receive static lean labels based on widely-documented editorial positioning. These are not AI-scored per article because the stance-toward-government axis is not meaningful for non-Bangladeshi coverage.
Story clustering
Multiple outlets covering the same event are grouped into a single story. The current clustering method uses slug similarity: stories with near-identical URL slugs (derived from headline text) are grouped together. This reliably catches near-duplicate headlines from the same event.
Cross-source clustering — grouping stories about the same event even when outlets use different headlines — is a known limitation at the current scale. When two outlets cover a protest by one using the word "demonstration" and another using "unrest," the slug-based approach may not group them. We plan to improve this with semantic similarity matching as the site scales.
Feed sources
Articles are fetched from public RSS feeds published by each outlet. The cron runs approximately every five minutes. What you see from each article is the headline, the outlet's name, the time it was published, the lead image if the outlet's feed carries one, and the first sentence or two from that feed. We do not republish full article text. All stories link directly to the original article on the outlet's own site.
Limitations and accuracy
AI lean scoring is not perfect. The model can misread irony, satire, opinion pieces presented as news, or articles where the framing is subtle. A score that looks wrong is always worth questioning.
Known failure modes:
- Opinion and editorial pieces may score as strongly leaning even when the outlet itself is independent
- Breaking news articles published before full context is known may score differently than follow-up coverage of the same event
- Bangla-language articles are analyzed by the same model with generally slightly lower confidence than English-language articles on the same topic
- Stories with few sources in the feed may show a lean bar dominated by one perspective simply because only one outlet has published yet
If a label is clearly wrong, let us know. We review flagged articles.
What is ours and what is not
The words in every article linked from this site belong to the outlet that wrote it. We do not rewrite headlines. We do not edit or add to any article.
Two things on this site are ours. The lean label next to each article, which is made by an AI and explained above. And the "Analyze coverage" note on a story page, which you have to ask for. That note is written by an AI from what the outlets have already published. It says what they agree on and where they differ. It can be wrong, and it is not a substitute for reading the articles. Every story links to the original on the outlet's own site.
The ownership label next to each outlet comes from public records. We show an accuracy rating for some international outlets. We do not show one for Bangladeshi outlets for now.