What’s the big deal?

Every week is full of big stories. This tool tries to answer, with data rather than instinct, which one mattered most. It measures every candidate story through five independent lenses, combines them two ways, and then tests how fragile each answer is.

1. Find candidates. Every event on Wikipedia’s Current events portal for the week, a daily, human-curated, sourced record of world news, is grouped into stories. The 20 most prominent stories (by days covered, number of events and number of cited outlets) are shortlisted. An AI pass merges duplicates.
2. Reach. How widely the world’s media covered the story: the number of world regions (of eight), source countries and languages publishing about it, from up to 250 articles per story in the GDELT news index. If GDELT is unavailable, reach falls back to how many Wikipedia language editions cover the story plus the outlets the portal cites.
3. Persistence. Whether the story lasted: days it appeared on the portal, plus days its daily GDELT coverage stayed above a quarter of its peak.
4. Attention. What people actually looked up: extra Wikipedia page views for the story’s article above its own four-week baseline, summed across 16 language editions, plus the number of languages where views surged.
5. Human stakes. How many people the story directly affects, estimated from the reporting (deaths weighted ten times), on a log scale so a million is not treated as a thousand times more than a thousand. Each estimate cites its basis.
6. Consequence. How much the story will matter in a year: a jury of three independent AI assessors scores scope, severity, irreversibility, duration and spillover (1–5 each). The spread between jurors is reported as uncertainty.

Combining the lenses. Raw measures are on different scales, so each lens becomes a percentile rank within the week’s shortlist. Two weightings then produce two answers:

• Impact: human stakes 35%, consequence 35%, reach, persistence and attention 10% each. What mattered most.
• Attention: attention 40%, reach 25%, persistence 20%, human stakes and consequence 7.5% each. What the world noticed most.

Stress test. Any fixed weighting is a judgement call. So each answer is re-run with 10,000 random weightings drawn around its profile (a Dirichlet distribution, concentration 20). A story’s win share is how often it still comes first. A high win share means the answer does not depend on the exact weights chosen.

What’s automated. Data collection, scoring, the stress test and this page are deterministic Python. An AI agent (Claude) handles only bounded judgements: merging duplicates, estimating human stakes with cited evidence, the consequence jury, and the plain-language write-up. Everything runs every Monday and the full data for each week is published below.

Every input and output for this week is public:

• candidates.json
• judgements.json
• result.json
• writeup.json

Limitations:

• Candidates come from English Wikipedia’s Current events portal, which has its own editorial biases toward some regions and topics.
• GDELT samples online news; coverage of some countries and languages is thin.
• Page views measure curiosity, not importance, which is exactly why attention is reported separately.
• Human stakes and consequence are AI estimates from reporting, not ground truth. Their evidence and juror spread are published.
• Long-running stories (a war in its third year) compete with sudden ones; persistence and the four-week baseline partly balance this.