In development, live now
Every major paranormal dataset plotted on one map, 385,531 reports, with a math engine that separates the signal from the noise.
Runs in a desktop browser with nothing to install. The first visit loads about 30 MB of data, then your browser keeps it.
01 / THE PROBLEM
Plot every report and you get a population map. The Signal Engine builds an expected count for each hex cell from census population and the region's reporting rate, then tests the real count against it.
A cell stays lit only if it beats that baseline: a Poisson test, Benjamini-Hochberg false-discovery control at q below 0.05, and a rate ratio of at least 1.5. Cells with significantly fewer reports than expected show in blue. Big cities alone do not light up.
02 / DEEP DIVE
Pick a point and the app runs a 50 km survey around it: an anomaly verdict, per-category rate ratios, space-time flap episodes, seasonality against the national pattern, and a report you can export as a standalone HTML file.
Sedona, Arizona logs 699 reports against 384.1 expected, 7.28 times its population baseline (p below 0.001). The dossier then argues against itself, listing nearby airports and other overlays that could explain the number.
03 / THE BORING EXPLANATIONS
Twelve overlay layers sit on the same map: military and DOE sites, restricted airspace, airports, national parks, earthquakes, caves, NASA fireballs, cryptid sightings, Missing 411 cases, historic sites, geomagnetic storms and a per-capita heatmap.
Here, the Nevada Test and Training Range with military sites, restricted airspace and airports switched on, one at a time. Spatial correlations report the raw r next to the population-adjusted r, so "ghosts correlate with UFOs" cannot hide "cities exist".
04 / THE ANALYST
SIGNAL is a built-in analyst powered by Claude. It drives the map through tool calls: hunting anomalies with the Signal Engine, running deep dives, filtering, correlating, placing pins, capturing the map into illustrated reports. Every call shows up in a live investigation log.
The app says you can try it with no setup while a free daily allowance lasts. After that, paste in your own Anthropic API key, which stays in your browser. All map features work without the analyst.
05 / HOW IT WAS MADE
Stewart Gregerson built Strange Signals with Claude Code as his pair engineer. It started on 13 March 2026 as a plain correlation map. In June a written diagnosis turned it into something stricter.
The v2 vision document opens by calling v1 out: it plotted where people report things, so population drove every number. Hexes were tested with no multiple-comparison control, and "UFOs correlate with Haunted Places" mostly meant "both happen where people live". The fix was a pure-math module, signal-engine.js, with no map code in it.
The app is plain HTML, CSS and JavaScript on Leaflet, Turf and D3, with no bundler. A Python pipeline merges nine source CSVs and geocodes about 109,000 NUFORC reports. The UFO file is split into three gzipped shards to stay under Cloudflare's 25 MiB per-asset limit, each parsed in its own Web Worker and cached by a Service Worker.
A rumored surge in NASA fireball detections was tested against primary CNEOS counts and refuted; the brief is in the repo. When a pipeline dedup bug was found truncating every description, the whole dataset was rebuilt rather than patched.
Strange Signals plots public datasets. Each belongs to its authors and keeps its own terms; check the source before reusing any of it.