Global hazard picture

Every marker is a live record pulled minutes ago from a public feed: USGS seismicity, GDACS multi-hazard alerts, and NASA EONET open events — plus the ✕ markers, this network's own seismic association detections, located from real-time station data rather than any catalog. Diamonds are the fusion engine's watch regions, colored by their current alert level. Scroll or drag to zoom and pan, click a hotspot ring or watch-region diamond to jump to its detail below, click any marker to pin its tooltip open, and click a legend entry to show or hide that layer.

Dynamic watch cells — wherever Earth is loud right now

Nobody curates this list. Seeds come from whatever the feeds report anywhere on the planet — GDACS Orange/Red alerts, quakes ≥ M5.5, storms and fire complexes — clustered at 400 km and ranked. The top cells then get the same hydrological enrichment as the curated corridors: discharge anomaly against that exact point's own 92-day baseline, plus forecast rain and any community reports nearby. A new disaster creates its own watch cell; no one has to edit a config.

Fusion board — mountain-hazard corridors

Each region fuses five independent signal families into one transparent score: hydrology (GloFAS discharge anomaly vs its own 92-day baseline), forecast rainfall, nearby seismicity, official GDACS alerts, and a static exposure prior. No single feed decides; the 26 August Langtang collapse was invisible to hydrological forecasting alone.

What the world's media is reporting

Media coverage lags the physics but leads the official assessments, and it is the only channel that reports impact — bridges gone, villages cut off. Headlines below are live from Google News RSS (keyless, verified); each watch region also runs its own place-name query, strictly filtered so generic hazard coverage doesn't inflate quiet regions. GDELT is wired in as an optional second backend. News feeds the fusion score as a corroboration signal, capped at 10 points.

Ground truth — the channel every satellite misses

User-submitted observations from people standing next to the hazard. No single report is trusted: a report's weight is its reporter's reputation multiplied by corroboration from other reporters nearby in time and space. One high-reputation reporter alone raises a flag; only a corroborated cluster moves the fusion score. You can file a report right here: it publishes a new version of this page's data with your viewer identity, every open view live-reloads to it, and it counts toward the nearest watch region on the next fusion run. Reports marked DEMO are a labeled replay of the 26 August Langtang timeline — they show the trust math, and being older than the 48-hour fusion window they contribute exactly zero to today's live scores.

Langtang 2026 — what warning was physically possible?

The wave now runs on the corridor's real terrain: an elevation profile fetched point by point down the Lende–Trishuli valley, driving a slope-based debris-flow law (v = K·√(S−μ)·e^(−x/L)) calibrated to the one hard published number — the first 22 km in ~7 minutes. Deceleration comes from the mountain, not from guesses, and a thousand Monte Carlo draws over friction and attenuation give every settlement an arrival band (shaded, p10–p90), not a single time. Drag the latency slider to see how the pipeline budget buys lives — leads are computed against the median wave; the band shows how much a fast flood steals.

Will the alert survive the disaster it warns about?

The Langtang flood destroyed the infrastructure an alert would have travelled over — the cell tower and fiber junction sat in the water's path. Three spores network-simulation packs (packs/disaster/, 18/18 assertions green) prove the topology claim: a centralized alert path goes dark below the cut while the wave is still travelling; a valley mesh of short independent hops with local alert authority keeps a path to every settlement; and at Mars light-lag the center physically cannot participate at all — local autonomy is the only architecture that scales off-planet. The Monte Carlo below puts population numbers on it — and each comms trial now rides one of the terrain model's flood draws, so a fast flood cuts the tower earlier and shortens every warning window in the same trial. Failure draws: sensor uplink, monsoon fiber outage, panic congestion, tower cut, per-hop radio death with storm-correlated multipliers (weather kills radios together, not independently), and skip-link redundancy on the mesh chain — the design a real deployment should specify, so the model rewards it.

AI situation brief

How it lands on OpenKO

Everything below the feeds already exists in the workspace — the early-warning system is a new organ grown from existing tissue, not a new organism.

Verified feed catalog

Every "live" row was fetched successfully from this machine while building this page.