Bangladesh Media Watch

Since August 5, 2024, the Awami League government collapsed following the quota reform uprising. BNP leads the current government; Jamaat-e-Islami and the newly formed NCP (National Citizens Party, born from the uprising) are the main parliamentary opposition. The lean scale is calibrated to this reality: −1.0 = strongly opposition-aligned (Jamaat/NCP/AL press) · +1.0 = strongly pro-BNP ruling government. Outlets formerly anti-AL are now pro-government; outlets formerly pro-AL are now in opposition or repositioning. Jamaat papers (Naya Diganta, Inqilab) are now coded as opposition to BNP, not government supporters.

Each article is scored using Moral Foundations Theory (MAC framework). Eight moral domains are measured: Deference to authority, Fairness, Group loyalty, Property rights, Islamic obligation, Heroism, Reciprocity, Family. Domain scores are precision-weighted and fused through a Bayesian-Kalman filter. Expert lean priors seed the posterior; real articles update it daily. The 95% conformal prediction interval reflects statistical confidence in each estimate.

Confidence = max(0, (2 − CI_width) / 2 × 100). A narrow band (many real articles processed) gives high confidence shown in green. A wide band means fewer real articles — the score relies more on the expert prior, shown in amber or red. Outlets blocked by access controls are marked "Low data." Adding real article text through the Data Input page immediately narrows the band and increases confidence.

A Bayesian Beta(α,β) posterior updated by two signal types: (1) article-level EPI scores measuring AI generation or epistemic manipulation, and (2) hard fact-checker verdicts from three IFCN-certified organisations. Recent verdicts are weighted more heavily with a 90-day exponential decay. Outlets accumulating debunked content see credibility fall automatically over time.

Estimates AI-generated or structurally manipulated content using an ensemble of signals: bigram repetition rates, lexical diversity, and AI self-identification patterns in Bengali and English. A multi-model ensemble combines XLM-RoBERTa, mDeBERTa, and structural heuristics. Higher EPI = more suspected AI-generated or templated content.

Scores are automated estimates across many articles — not human editorial judgement. An outlet can produce excellent individual journalism and still show consistent framing patterns across its total output. Political contexts shift rapidly; priors are updated when editorial repositioning is documented. Treat scores as quantitative evidence for critical reading, not a verdict on any outlet or journalist.