Every article is scored against seven moral and political frames drawn from Moral Foundations Theory and adapted for Bangladesh's political context. Each frame measures how strongly an article favours or opposes the current ruling government (BNP) versus the parliamentary opposition (Jamaat-e-Islami / NCP).

The seven frames are: Deference (respect for authority and government), Fairness (impartial treatment and rule of law), Group loyalty (in-group solidarity and nationalism), Property rights (economic freedom and ownership), Reciprocity (accountability and justice), Family values (traditional social structures), and Islamic framing (references to religion in political context).

Article scores are combined with an expert prior for each outlet and updated using a Bayesian formula. The result is a lean score on a scale from −1.0 (strongly opposition-aligned) to +1.0 (strongly pro-BNP government). A 95% confidence interval shows how reliable each outlet's score is — narrow bars mean many articles were scored, wide bars mean the score relies more on the expert prior.

Each article receives an Epistemic Integrity (EPI) score measuring the likelihood that it was written by AI or contains AI-amplified disinformation signals. The score looks for statistical patterns that distinguish AI-generated text from human journalism: unusual word repetition, lack of source attribution, hyperbolic language, and structural regularity.

The outlet-level credibility score (0–100) is a Bayesian posterior combining the daily EPI scores with a baseline prior derived from each outlet's historical track record and IFCN fact-checker assessments. Higher scores indicate more reliable, human-authored journalism with diverse sourcing.

IFCN-certified fact-checkers (Rumor Scanner, Dismislab, AFP Bangladesh) are treated separately — their lean score is fixed at 0.00 by design, as they are structurally independent of political alignment.

Articles are collected daily at 08:00 Bangladesh time (02:00 UTC) via targeted Google News queries for each of the 25 monitored outlets, supplemented by direct RSS feeds where available. Approximately 2,300–2,500 articles are scored each day.

Only articles containing substantive political content are scored — articles about sports, weather, entertainment, or purely commercial topics are filtered out before scoring. Each article must pass a political relevance threshold before entering the lean scoring pipeline.

The platform analyses images, audio clips, videos, and political photocards for signs of AI generation or manipulation. Image analysis uses frequency-domain forensics to detect GAN and diffusion model artifacts. Audio analysis detects voice cloning by measuring pitch regularity and spectral characteristics. Political photocards are analysed for outlet colour signature mismatches and AI-generated text patterns.

All detection results are accompanied by a calibrated uncertainty estimate. Items flagged as suspicious are added to a priority queue for review by IFCN-certified fact-checkers. The system improves over time as verified fake content is added to the training dataset.