Each article is scored using real sentiment analysis — not just keyword counting. The system reads whether the article is praising or criticising the government. It uses 80+ politically valenced phrases in Bengali and English, detects grammatical negation to flip signals when appropriate, and gives extra weight to article headlines (which carry stronger editorial intent). Scores from all articles that day are combined using a Bayesian precision-weighted average, then refined over time via a Kalman filter that tracks how each outlet's editorial position evolves. The expert prior — calibrated by a human researcher — seeds the model but real daily articles progressively update it.

The scoring uses Moral Action Coding (Hopp et al. 2021): eight moral domains — Deference to authority, Fairness, Group loyalty, Property, Islamic obligation, Heroism, Reciprocity, Family — are scored from article text and projected onto the political axis. In the Bangladeshi context, deference framing and group solidarity tend to correlate with pro-government content; fairness and rights-violation language tends to correlate with opposition framing.

Each outlet receives a credibility score from 0 to 100. It is updated daily by two signals: the volume of AI-generated or manipulated content detected in the outlet's articles (which lowers the score), and verified verdicts from three independent IFCN-certified fact-checkers (which raise or lower it depending on what they find). Recent verdicts carry more weight than old ones — the system "forgets" slowly over 90 days, so outlets can recover credibility if they improve.

The EPI measures how much of an outlet's content appears to be AI-generated, templated, or epistemically manipulative. It detects repetitive sentence structures, low lexical diversity, and explicit AI self-identification signals in both Bengali and English. Higher EPI means more suspected synthetic or coordinated content.

Every lean score comes with a 95% confidence interval shown as a horizontal range bar. A wide band means fewer real articles were processed that day — the score leans more on the expert prior. A narrow band means many real articles were scored — high empirical confidence. Outlets marked "Low data" depend more on the prior; add articles via Data Input to improve their accuracy.