The model (transparent Bayes): we start with a prior P(credible) from the source's reputation, then update it with evidence as likelihood ratios — substantive detail and specificity push it up; clickbait/sensational language and thin sourcing push it down. The result is a posterior probability, shown with its math on every card. v2 uses the signals available from the feed; the full model adds the 10 Tells (multi-source corroboration, primary data, named officials, chain-of-custody, hoax-pattern, persistence, reproducibility) as additional evidence — and calibrates against real outcomes.