npm run demo runs it locally against the live feeds.
StatusExperiment, local demo, hosted deployment paused
RuntimeTypeScript, Cloudflare Worker and D1 via Wrangler, React and MapLibre
Try it
npm ci && npm run demo, no API keysProof hooks315 offline tests, reproducible severity-gate evaluation
How it works
scripts/demo.ts creates a temporary directory with random owner, collector and worker tokens. It applies the D1 migrations to a throwaway local database and starts wrangler dev on port 8797. It then runs the collector for four registry entries: global-usgs-earthquakes, global-gdacs-alerts, us-nws-warnings and ca-eccc-warnings. A preview server opens on http://127.0.0.1:4319/?hours=168, the collector reruns every five minutes, and Ctrl-C deletes the temporary directory. The demo calls no model.
collectSources fetches each source from its registry URL with a byte cap and hands the response to that source’s parser. Each parser turns issuer records into observations that carry the canonical record URL, the issuer’s fields as a JSON excerpt and an official block with activity and impact. The server’s ingest in src/server/index.ts publishes an official record with a location as an incident straight away, with no model step. A changed record becomes a new revision of the same incident.
Impact comes from the issuer’s own level, mapped one way per source, and never reaches critical:
NWS Minor maps to
informational. The source notes explain why each level maps down a step: PAGER and GDACS levels are modelled loss estimates and CAP severity states a threat, so none of them is reported harm. Most earthquakes show unknown because USGS issues PAGER alerts only for damaging quakes. GDACS also lists archived events among its Orange and Red results, so the collector publishes only events GDACS marks as current. After each complete read it sends the current IDs to /api/sources/<id>/reconcile, which ends incidents GDACS no longer lists.
Narrative sources such as news and police logs take a second path. A local Codex worker (src/worker/adapter.ts, requesting gpt-6-luna) extracts claims, and the host decides whether each claim may stand. validateSeverity rejects a severe or critical claim unless it quotes an affirmed consequence or an official instruction from its own cited excerpts. Negated, hypothetical and drill wording is rejected using per-language lexicons, and Chinese, Japanese and Thai use substring matching. Until a worker connects, the brief panel shows a model-free brief from structured-brief.ts built from issuer fields.
The thermal layer in src/server/geospatial.ts reads FIRMS’s public 24-hour global VIIRS CSVs for NOAA-20, NOAA-21 and Suomi NPP, which need no MAP_KEY. It filters to the visible area on the server and caps a view at 2,000 detections, sampling across a 48 by 24 grid so one busy region can’t crowd out the rest.
Sources and terms
Map tiles come from OpenFreeMap with OpenStreetMap attribution. Place names come from a GeoNames cities15000 extract (CC BY 4.0) and borders from Natural Earth 1:110m (public domain), both bundled, so there are no geocoding calls. The National Hurricane Center is not a source. Tropical cyclones arrive through GDACS and as ECCC hurricane warnings. NOAA tsunami feeds and USGS volcano alert levels were reviewed and left out, with reasons in
docs/sources.md.
The application code is UNLICENSED, so a public repository grants no reuse rights.
Severity gate evaluation
docs/severity-evaluation.md runs the real validateSeverity against Wikipedia lead excerpts about past earthquakes, labelled severe when the event’s USGS PAGER alert is orange or red. The dataset, eval/severity/dataset.json, holds 1,908 excerpts about 264 events (76 severe) in 22 languages. Excerpts are pinned by revision ID under CC BY-SA 4.0, PAGER labels are public domain and the Wikidata links are CC0. No model is called. Two deterministic proposers stand in for one: eager always proposes severe and quotes the first sentence, and lexical-alarm escalates on keywords without checking polarity.
At commit c8967c7, with 95% bootstrap intervals over whole events:
The gate removes escalations whose quote doesn’t state a consequence. It barely changes agreement with PAGER for the keyword proposer, whose precision moves from 51.1% to 51.8%. Much of the residual is a label mismatch, because PAGER models loss while the excerpts often report real deaths. The write-up also records a design gap. The gate checks that a quoted sentence is in the excerpt and affirmed, but not that it describes the same event, and a sentence about an 1856 quake passed for a 2021 one. The evaluation covers earthquakes only, uses retrospective encyclopedia text rather than live news, and measures no real model. Results are in
eval/severity/results.json.
Verify it
These checks run offline:c8967c7: npm run check passed, npm test reported tests 315, pass 314, fail 0, skipped 1 (the opt-in test that needs an authorised Codex host), npm run build passed and npm run audit:public printed Public-source guard passed for 187 tracked files. npm run eval:severity printed the numbers above and rewrote results.json byte for byte.
To see the live map:
AGENTS.md says agents must not start collection while the project is paused, and the demo starts a local collector against live feeds. It needs Node.js 24 and ports 8797 and 4319 free.
Limits
- The hosted deployment was paused on 21 October 2026 (
docs/paused.md). There is no public URL, and the restart checklist is unfinished. - USGS and GDACS are the only worldwide incident sources. NWS covers the US and its territories and ECCC covers Canada. NWS zone-only alerts (winter, heat, tropical, marine) and ECCC watches and advisories are not collected.
- Map points are epicentres, alert centroids or polygon bounding-box centres with a radius, never impact footprints.
- Thermal detections link to the FIRMS map, not to a per-detection record, and they don’t establish incidents or severity.
- GDACS liveness depends on the collector running. A stopped collector leaves the last current set on the map.
- Narrative claims and the five-minute model brief need a local Codex worker with its own sign-in. The demo doesn’t start one.
- The severity gate has no event attribution and has been measured on earthquakes only, with stand-in proposers.
- No endurance run, restore drill or capacity measurement has been done.
- Pull request #3 proposes more sources and a public demo. It was open and unmerged on 23 October 2026, and this page describes
main.
results.json has data for 22 (Swahili matched no article). It asks for Node.js 24, while package.json declares >=22.16. Its tagline promises every marker linked to its public record, which holds for incidents but not for thermal detections.