The map
Gold dots are the fossil records the two models learned from: west of the Cascades (left of -121.5°) and east of them. Grey ones in the east are small animals, birds and fish rather than big mammals; hollow ones had no terrain values at all. The targets are coloured by what the re-check found under them. Tap anything.
How far to trust it
I built IceWave in early 2026, a sequel to PaleoWave, and it looked like a success: high scores, a "blind validation" at a real mammoth dig, a 90% leave-one-out recall. Coming back to it this fall, I re-checked every headline number from the files in the repository. None of them survives. The arithmetic is in tools/check.py, so anyone can rerun it.
Where to look instead
Washington's best-known Ice Age finds have one thing in common: someone was digging. Quarry workers found the Coyote Canyon mammoth in a rock pit south of Kennewick in 1999. Emanuel Manis found the Sequim mastodon with a backhoe, digging a pond, in 1977. A road crew found the Wenas Creek mammoth near Selah in 2005, and builders found a mammoth tusk 30 feet down in a Seattle apartment excavation in 2014. The ground that holds them is Ice Age ground: flood deposits, loess, old lake beds and bogs, glacial gravels. Bedrock basalt doesn't.
So IceWave now maps every active surface-mine permit in Washington and checks what each pit is cut into, on the state's 1:100,000 geologic map. Every Monday it checks the permit list again, so a new pit in Ice Age ground shows up here.
The biggest pits in eastern Washington's Ice Age ground
| Pit | Operator | County | Cut into | Acres | Depth |
|---|
The old targets
The model scored terrain cells across five states, kept the best cell in each one-degree square, and listed the top ones: 17 west of the Cascades and 25 east. Here's each one with what's actually there. The final east model (v4) re-ranked the same 25; the column shows where each one ended up.
"Level patch" is the area around the target where the 3DEP elevation surface is flat to within a centimetre. Elevation models draw lakes, reservoirs and wide rivers as one flat level, so a patch of a square kilometre is water. Water names come from OpenStreetMap. The full lists are icewave_v3_top50_lidar.csv and icewave_v4_final_top50.csv.
Close-ups
For twenty east targets I pulled USGS 3DEP elevation at about 14 m and drew a hillshade, the topographic position and slope. One more correction: the middle panel is labelled "blue = valley", but its colours run the other way, so red is low ground and blue is high. The README called these LiDAR; they're 3DEP elevation fetched at 14 m, which in eastern Washington is built from lidar but isn't raw lidar. Look at the star: on many of them it sits on a perfectly smooth, featureless surface. That's water.
The real mammoth
The Coyote Canyon Mammoth Site is real and well worth a visit. Quarry workers south of Kennewick found big bones in 1999; the McBones Research Center has been excavating a Columbian mammoth there since 2008, in deposits left by the Ice Age floods that poured out of Glacial Lake Missoula and across eastern Washington between roughly 18,000 and 15,000 years ago. They run tours.
It's a fine example of exactly what IceWave was hoping to find: big animals caught and buried by the floods in the Columbia Basin. It just isn't where the model pointed.
Learn more, and book a tour, at mcbones.org. It's an active, permitted excavation on private land; please don't visit outside a tour.
If you go
Look, photograph, report. Don't collect.
Vertebrate fossils on federal land can only be collected under a permit, under the Paleontological Resources Preservation Act. Much of the Columbia Basin, the Willamette Valley and Puget Sound is private land: farm fields, quarries and back yards, so ask before you walk. If you find bone, leave it where it is, photograph it with something for scale, note the coordinates, and tell the landowner or land agency, or a museum such as the Burke in Seattle.
And skip the targets on water. Several of them are in the middle of a river or a reservoir.
What's next
The pit list is the useful part now: a short, real list of places where Ice Age ground is being opened up this year. The model itself still needs rebuilding, not tuning:
- Measure every input the same way at fossil records and at background points: the same elevation source, and real geology at both, not a default for one side.
- Mask out water before scoring, so a flattened river can never be a target.
- Train on the animals it's meant to find (large mammals), or model small animals separately; and drop records whose coordinates are only a county centroid.
- Test by leaving out whole places, as the re-check does, and report that number.
- Score the whole landscape, then compare the top cells with places like Coyote Canyon that the model never saw.
Meanwhile the weekly checks keep running: PBDB for new Ice Age finds, and the state's permit list for new pits in Ice Age ground. If you work at one, or live near one: bone in a gravel face is worth a phone call to the Burke Museum.