About EuroPrisma

A pipeline runs twice daily (07:00 and 17:00 Europe/Berlin). Claude reads the day's European news and proposes 3–5 events worth surfacing. Crucially, the AI doesn't write the event itself - for every proposal, the pipeline searches Event Registry, an external media database that indexes hundreds of outlets across Europe, and asks Claude to pick the matching real-world story cluster (or reject it if no good match exists). Only proposals tied to a real, well-covered Event Registry cluster make it onto the site. So even though the proposing step is AI-driven, the events you see on EuroPrisma are real news clusters with real article counts behind them - not anything the AI made up.

For each published event, we ask Event Registry how many of our seeded outlets per country wrote about it in a 7-day window. We then display, for every country, the share of its seeded outlets that covered the event - for example, "10 of 17 Italian outlets covered this event" shown as 59%. We use a percentage of the seeded list rather than raw article numbers for two reasons:

• Raw counts mislead across countries. Germany's media market is many times the size of Slovenia's, so 50 articles in Germany and 5 in Slovenia can represent the same level of public attention.
• We legally cannot store article text - only metadata (URLs, outlet, publish date, concepts mentioned). The percentage is what we can measure consistently from that metadata.

Hover any country to see the underlying counts (e.g. "3 of 17 outlets").

Each country has a curated outlet list - typically 15 to 20 outlets, scaling from 7 in the smallest countries up to 25 in the largest. The list isn't editorial taste; outlets are chosen against three criteria:

• Reach threshold - outlets must reach at least ~17% of the country's online news audience, verified against primary-source web traffic data (Similarweb / HypeStat) rather than self-reported numbers.
• Spectrum balance - left-leaning, centrist, and right-leaning outlets are represented per country.
• Format diversity - print/digital, broadcaster sites, and news agencies where applicable.

A small number of outlets are excluded from the count when their format makes article numbers misleading - for example liveticker-heavy sites that publish dozens of fragments per event. The full list, plus the per-country research that produced it, lives in the project repository.

A single one-dimensional spectrum the event might divide Europeans on. The AI proposes a left pole and a right pole (e.g. Sovereignty ↔ Solidarity, or Caution ↔ Action), with neutral language at both ends. The axis isn't trying to capture the whole event - just a dividing line.

If the AI can't produce a clean axis with both poles and a description, the event is dropped from that pipeline run and never published.

All metrics on the Insights page are derived from votes:

• Country average - for each event, we average the vote positions (–1 to +1) of voters from a given country.
• Country alignment - distance between two countries' average positions, averaged across events both countries voted on.
• Divisive events - events where country averages are spread far apart.
• Consensual events - events where country averages cluster together.

Events that have only been voted on within a single country don't appear in the divisive or consensual lists - there's no cross-country comparison to make yet.

Event titles and summaries are produced in all 12 supported locales (English, German, French, Spanish, Italian, Polish, Dutch, Portuguese, Romanian, Czech, Hungarian, Greek) at pipeline time. User comments stay in the language they were written in, with a per-comment Translate button. English is the fallback when a translation is missing.