Triassic Β· Nevada Β· Rechecked Oct 2026
Ichthyosaurs Β· Random Forest Β· PBDB Β· USGS 3DEP Β· Live page
π Project PaleoWave
Two hundred and forty million years ago, Nevada was the bottom of a warm shallow sea.
A random forest that learns the terrain where central Nevada's Triassic ichthyosaurs have been found (elevation, slope, ruggedness, whether a spot sits in a basin or on a ridge) and points to fifty places that look alike. Version 3 splits the north (Prida and Favret) from the south (Luning and Gabbs).
Rechecked in October 2026: the 30 training records sit at only 18 places, so the old leave-one-out scores were flattered; held out place by place, the model recognises 7 of 16. It still tells fossil terrain from random terrain well. The new page shows the map, the homework, and a weekly check of the Paleobiology Database. Version 4 searches only the mapped Triassic marine formations, finds the bare limestone in Sentinel-2 imagery, and lists forty places to look on BLM land.
Ice Age animals Β· Random Forest Β· USGS 3DEP Β· Live PBDB
𦣠Project IceWave
I built IceWave to find where the Missoula Floods buried mammoths. Rechecked, its favourite targets were rivers and reservoirs, and its "blind validation" was 69 km off. The page shows how that happened.
Random forests trained on Pleistocene fossil records across Washington, Oregon, Nevada, Idaho and Montana, one west of the Cascades and one east. The October 2026 re-check found two shortcuts: 23 of the 40 east records had all-zero terrain because the elevation grid never reached Montana, so zero slope (water) looked like fossils; and every background point got the same default rock score, so "not the default" meant "fossil".
Tested fairly, leaving out whole places and without the rock score, the east model scores an AUC of 0.65, where a coin toss is 0.5. The page colours every target by what's actually there and checks the Paleobiology Database every Monday. And since Washington's Ice Age finds all came from someone digging, it now maps the state's 478 active pits cut into Ice Age ground, and flags new ones weekly.
The model didn't call its shot
This box used to say IceWave predicted the Coyote Canyon mammoth dig "within miles of the predicted pixel": three methods, one answer. That was wrong, and I want the correction to sit where the claim did.
E02 is at 46.30°N, 120.13°W in the Yakima Valley, on a patch the elevation model has flattened to a single level, the way it draws ponds and rivers. The Coyote Canyon Mammoth Site is south of Kennewick, 69 km away. The bones were found in 1999 and the McBones Research Center has dug there since 2008, so it wasn't blind. The "validated" label came from a line of code that marked whichever target was first. The final model ranks E02 23rd of 25.
| Claim | Was | Rechecked |
|---|---|---|
| E02 to the dig | "the predicted pixel" | 69 km |
| East leave-one-out | 36/40 (90%) | AUC 0.65 without the default rock score, whole places held out |
| West AUC | 0.890 | 0.63 on terrain, whole places held out |
| "Valley floor" targets | 13 of 20 | 12, and 8 of those are water |
The real mammoth is well worth a visit: McBones runs tours. The arithmetic is in tools/check.py.
π PaleoWave β Top 10 (Nevada)
𦣠IceWave β Top East Targets, Rechecked
PaleoWave β Triassic Nevada
240 million years ago, Nevada was the bottom of a warm shallow sea. Giant marine reptiles β ichthyosaurs, the biggest of them as long as a bus β cruised these waters, hunted, died, and sank into the carbonate mud. Then a continent happened.
Basin and Range tectonics folded and faulted the seafloor into the desert mountain ranges we drive past on I-80. Erosion has been working on those limestone layers ever since, slowly uncovering the bones. Berlin-Ichthyosaur State Park found some of them. The Triassic marine formation covers hundreds of square kilometers of central Nevada outback β largely unwalked by paleontologists.
The key insight: fossils don't appear at random. They appear where Triassic limestone is being actively exhumed β basins and lower slopes where erosion strips the overburden. An honest caveat, found in the October 2026 recheck: the model knows only 18 places, and its earlier "strongest predictor" was partly an artifact of measuring fossil sites and comparison points two different ways. The page shows what still holds, and the recheck post tells how I found it.
IceWave β Pleistocene West
The Missoula Floods were among the largest floods in Earth's history β repeated catastrophic releases from Glacial Lake Missoula that scoured the Columbia Basin and deposited slack-water sediments across eastern Washington and Oregon. Megafauna that drowned in these events were buried in exactly those slack-water deposits.
IceWave set out to find Columbian mammoth, mastodon, horse, camel, bison, ground sloth and short-faced bear, with separate random forests for the wet west (Willamette Valley, Puget Sound) and the dry east (Columbia Basin, Owyhee, Basin and Range), split at -121.5°.
An honest caveat, found in the October 2026 recheck: the east model learned from 40 records, 16 of them small animals, birds and fish, and 23 with terrain values of zero; and it learned that a default rock score meant "not a fossil". Its top targets were reservoirs and lakes. The page shows each target as it really is, and what a rebuild needs.