Home › Methodology: confidence score and merged alerts
Methodology
Everything on alerts.live is descriptive. The confidence score says how well an alert is corroborated, not how dangerous a place is, and no page gives a safety score.
Confidence score (0-100)
Each alert starts at 10 points (8 for a visitor report) and gains points from independent evidence. Every factor is stored and shown with the alert.
| Evidence | Points |
|---|---|
| One independent news outlet | +15 |
| Each further independent outlet (2nd, 3rd, ...) | +8, +8, +8, +6, +5, +4, +3, +2, +1: diminishing returns, about 60 at ten outlets, outlets capped at 62 |
| A trusted wire or official source | 45, plus 8 per extra trusted source (up to 2), 4 per other outlet (up to 2), 2 per social post (up to 2) |
| An official open-data record (for example a road authority) | +65 |
| Visitor confirmations from distinct devices and networks | 12, 26, 42, 50, 56 for 1 to 5 or more; a burst within seconds counts less |
| Syndicated copies of one wire story | count once |
| Social posts | weigh less than outlets; one post alone stays low |
An uncorroborated visitor report is capped at 25 and a single confirmation can never reach the public threshold. Visitor reports appear publicly at 50 or more; news and official alerts are always published. The reporter's own device and network never count as confirmations, and the history of a reporter can lower or raise the score.
Verified
An alert is labelled verified when it comes from a trusted source, or has three or more independent sources and confirmations, or reaches a confidence of 50 or more. Unverified alerts are still shown, with a lower confidence.
Merged alerts
New headlines are compared with active alerts using a match score: compatible type, distance, time, and distinctive words and place names shared between headlines. Above 0.5 they are treated as the same incident: the sources are combined on one alert and the duplicate is hidden, so an incident counts once on every page, digest and count. Merges are logged and can be undone. Alerts that were wrongly grouped are split again by the same rule.
Danger indicator
A separate danger number (1-100) is estimated from the alert type and wording. It describes the event, not the place, and is not used for the confidence score.
Counts and snapshots
An alert is counted in a window (24 hours, 7 days, 30 days) when its latest sighting falls inside it. The 30-day chart groups alerts by the UTC day of their latest sighting. Zones leave the live map 12 hours after the last sighting; the history is kept. Pages with fewer than 2 alerts (cities) or 3 (countries) are marked noindex so search engines only see pages with real data.
Limits
- Coverage depends on what public sources publish and on automatic reading of headlines in many languages; mistakes happen (wrong place, wrong type).
- Absence of alerts does not mean absence of events.
- Always follow local authorities and official guidance.