An importance score, not a probability.
The score answers “should this be prominent for this user now?” It does not predict whether an event will occur.
1. Candidate generation
Independent adapters produce candidate signals from weather forecasts, NASA natural events, USGS earthquakes, public holidays, local sky calculations and selected sports fixtures. A failed adapter does not fabricate candidates.
2. Source-specific evidence
Weather uses source-supported thresholds. Earthquakes combine magnitude, distance and age. Natural events combine category, distance and recency. Sky events use annual peak dates plus lunar context. Holidays use date proximity. Sports uses national-team relevance and fixture timing.
3. Common ranking
4. Publish threshold and dedupe
Most low-scoring candidates are removed. Similar cards are deduplicated and the public feed is capped so the product does not turn into a noisy directory. Hazard-like signals can remain visible regardless of optional category preferences.
Limitations
Launch weather scoring uses absolute forecast thresholds rather than claiming record-breaking anomalies without a climatology comparison. EONET is curated, not an exhaustive local incident feed. Sports coverage varies by API and free-plan coverage. Approximate Cloudflare location can be wrong, especially behind VPNs or mobile networks.