Disparities, targeting patterns, and lead–lag structure across every documented case — with the methods stated and the receipts one click away.
How defendants from ethnic-minority regions fare against Slavic-majority defendants — sentence length, torture rate, and hate-crime context across the registry.
“Minority” aggregates defendants from designated ethnic-minority regions; comparisons are registry-wide and not charge-matched, so read disparities as directional, not causal.
Documented cases grouped by what people did for a living — filter by country to see how the pattern of persecution shifts across borders.
Occupation is recorded as free text and folded into curated buckets by keyword; cases that don’t match any bucket (≈5%) are excluded, so percentages are of classified cases. Journalist coverage is inflated by press-freedom sources in the registry — read the mix, not just the top bar.
Where spikes in repression-news coverage statistically precede arrest waves — and by how long. When one of these signals is elevated right now, the historical lead time becomes a concrete watch window.
Correlational lead–lag evidence, not causation. Only pairs passing a permutation test that preserves each series’ autocorrelation are shown. The live-watch banner fires when a signal’s recent complete weeks run ≥1.5σ above its own long-run weekly volume.
The dominant themes the model finds across thousands of case narratives — a map of what people are actually being prosecuted for, learned from the text itself.
Themes are unsupervised clusters of semantically similar case text (BERTopic); the outlier bucket is excluded. Counts are documents assigned to each theme.
Prophet time-series projection of arrest volume, with an 80% confidence band. The shaded region right of the marker is the forecast; everything left is the model fit to observed history.
Forecasts are statistical projections from historical arrest records and carry uncertainty — read the band, not the line. The upgraded model is event-aware (mobilization, elections, wartime censorship laws) with an 80% band conformal-calibrated against held-out backtests. Selections with sparse data fall back to the all-jurisdictions model.
Open the full interactive graph of cases linked by shared charges, locations, and timing.