Critical-mineral targeting · U.S. federal land
Atlas ranks open federal ground from public geoscience data, re-checks every target’s land status against federal records, and publishes how the ranking does on deposits it never saw, weak results included.
28 packages·10 states·622 open targets after the land re-check·run data as of 2026-09-22
Pick a district below to see it up close
Outlines mark the areas we analysed, never a target. Imagery: NASA Blue Marble via NASA GIBS (globe) · aerial imagery: USDA NAIP via USGS The National Map · relief: USGS The National Map · Sentinel-2: contains modified Copernicus Sentinel data (ESA) via Microsoft Planetary Computer.
What it is
For each district: a ranked short list of open federal ground, the plain-language reasons behind every target, and a public record of how the ranking has done.
At the run, every target had no active mining claim in BLM’s records (MLRS) and sat on BLM or Forest Service land outside BLM’s protected areas. A fresh re-check on 2026-09-30 added two checks the run lacked, Forest Service wilderness and recorded mineral withdrawals, and flagged 190 of the 812. Those are no longer offered as open ground.
Run data as of 2026-09-22.
28 packages across 10 western states: copper, gold, silver, cobalt, lithium, molybdenum, rare earths and nickel. Each target comes with the plain-language reasons it ranks where it does.
Every package shows its held-out test result, weak ones included, or says plainly that it has too few known producers to test. The target list is hash-committed so later changes would show.
Track record
We hide known deposits, re-rank, and count how many land in our top-ranked 20% of ground. Then we check that result against chance done properly.
4 of 8
district windows beat a strict chance test
Pass: Morenci, Tonopah, Ely and Carlin. Not yet: Creede, Bodie, Comstock and Silver Valley. The test keeps how deposits cluster and repeats our exact procedure (p < 0.05). These windows sit on well-known camps, so this measures ranking within a camp, not finding new districts.
2 of 8
statewide packages beat a strict chance test
Pass: Utah and New Mexico (the rows below show which commodity). Same strict test, which includes beating a no-geology rule that simply ranks known mining corridors first (p < 0.05).
47%
median share of known producers found in our top-ranked 20% of candidate ground
Held-out test across all 22 packages with at least 15 known producer records (range 36%–92%). Chance is not 20% here: it depends on how each package’s ground was chosen and tested, so each package’s matched level is the tick mark below.
Stricter view: the 16 packages we can re-check exactly from today’s files have a median of 54% per producer record, and 44% when each deposit cluster counts once.
622 of 812
listed targets still pass every land check on a fresh re-query on 2026-09-30
We re-ask the agencies’ own records: BLM mining claims, land agency, protected areas and withdrawals, plus Forest Service wilderness. A target counts only if every check answers and none rules it out; the rest are excluded and no longer offered. As a control, 126 of 126 sampled claimed cells are still claimed. This is agreement with public records, not a title opinion.
All 22 packages with at least 15 known producer records, strongest first. The weak ones stay in.
Share = known producer records (from the USGS mineral-deposit database) whose nearest cell ranked in our top 20%, with those deposits held out while ranking. “Not measurable” means we could not re-derive that package’s shipped ranking exactly from the files on disk, so we make no strict-test claim for it.
On 2026-09-29 we hash-committed all 812 listed targets, together with the claimed and withdrawn ground they are compared against (Merkle root 01cf92729b45…). Once we reveal a target, anyone can check it against that root, so the list can’t be quietly changed. Targets the later land re-check excluded stay in this record, so the exclusion is visible too.
External time-stamp: anchored in the Bitcoin blockchain (block 969224), dated 2026-09-29. Proof files on request.
Do new mining claims get staked near the pre-registered targets as time passes? This is the one outside signal we can get without a drill. Next check: 2026-11-01.
Does the ranking find more than a “look near known mines” rule would? We scored every cell of 22 state-and-commodity regions the model was not trained on. Atlas’s geology model, run without the “near known occurrences” signal as the plan required, put 60% of their known producers in its top-ranked 20% of ground; chance is 20%. A rule with no geology, which ranks ground by how many known mineral occurrences sit nearby, put 86% there and beat the geology model in 22 of the 22. That is a fail under the rules we time-stamped before any score was computed, and we publish it as one.
Held-out test. Known producers are split into spatial blocks. Each block is hidden in turn, the ranking is rebuilt without it, and we record where the hidden producers fall.
Candidate ground. In each package we rank a set of cells: some near known occurrences, some far from any. “Top-ranked 20%” means the top of that set, not of all land.
Strict chance test. Fake deposit clusters with the real cluster sizes are dropped on random cells and pushed through the same procedure, and a no-geology rule ranks known mining corridors first. A package passes only if its real result beats all of them (p < 0.05).
Ranges. Deposits cluster, so we resample whole clusters rather than single records. That gives honestly wide ranges; where a package has few clusters, they can still run slightly narrow.
One caveat. Some model settings were chosen while looking at these same districts, so the results are held-out for the deposits, not for every modelling choice.
Run data as of 2026-09-22; strict tests computed 2026-09-29.
Roadmap
The plan is to prove the ranking honestly, in public, before selling it.
2026 · done
28 packages ranked from public data and tested against known deposits wherever there are enough to test, weak results included. Every target’s land status was then re-checked against federal records.
September 2026 · done
All 812 listed targets time-stamped in the Bitcoin blockchain before any outcome is known, and a forward test started: do new mining claims get staked near them?
October 2026 · done
Scored on every cell of regions the model never saw, under rules time-stamped first. A no-geology “look near known mines” rule beat Atlas’s geology model in 22 of 22. We said we would publish the result whatever it showed, and we have.
Now
The open question is whether geology improves on “look near known mines” inside known districts, where the shipped rankings work. Alongside it: recruiting an economic-geology advisor and seeking research funding for field work.
Next
A hired field team samples Atlas's top targets, ground the “known mines nearby” rule would pick, and random control sites, without knowing which is which. A certified lab tests the rocks. That shows whether Atlas adds anything a simple rule doesn't.
2027
The forward test is scored at six and twelve months, in March and September 2027. With field results in hand, the first pilots with exploration teams follow.
How it works
Geology and fault maps, stream-sediment and rock geochemistry, gravity, magnetics, radiometrics and satellite alteration mapping. All public U.S. government and satellite data.
We split each area into cells and score each one on how closely its geology matches places where that commodity has been produced, and on how near it is to known occurrences.
Known producers are hidden in blocks and the ranking is rebuilt without them, then checked against chance.
Every target is checked against BLM mining-claim and land-agency records, protected areas, Forest Service wilderness and recorded withdrawals. Anything a check can’t confirm is left out.
The output: each district’s targets, ranked, with the evidence for each, their land status at the latest check, and the district’s tested hit rate, or a plain “too few known deposits to test” where that’s the truth.
Who’s behind it
“The U.S. needs domestic critical minerals, a lot of open federal ground is still unstaked, and the public data to rank it already exists. I built Atlas to turn that data into a short list you can check, and I publish how it performs, including where it doesn’t.”
Matthew Schatz, founder. Atlas Mining Systems is operated by The Atlas Project AI, LLC, a Florida company.
Contact
I’m looking for an economic-geology advisor, research partners for field validation, and early backers. Exploration teams interested in a future pilot are welcome to write too.
contact@atlasminingsystems.comcopper · district window · statewide
38 targets ranked
Held-out test: 84% of 63 known producer records fell in our top-ranked 20% of ground.
Chance test: beats strict chance (p = 0.0085; matched chance level 32%) — —.
Likely range 44%–97% (95% interval, resampling its 10 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
7 targets ranked
Held-out test: 66% of 44 known producer records fell in our top-ranked 20% of ground.
Chance test: beats strict chance (p = 0.029; matched chance level 32%) — —.
Likely range 32%–93% (95% interval, resampling its 14 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
13 targets ranked
Held-out test: 83% of 177 known producer records fell in our top-ranked 20% of ground.
Chance test: beats strict chance (p = 0.001; matched chance level 21%) — —.
Likely range 32%–96% (95% interval, resampling its 20 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
copper · district window · statewide
34 targets ranked
Held-out test: 70% of 70 known producer records fell in our top-ranked 20% of ground.
Chance test: beats strict chance (p = 0.037; matched chance level 30%) — —.
Likely range 12%–94% (95% interval, resampling its 10 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
silver · district window · statewide
24 targets ranked
Held-out test: 92% of 103 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.066; matched chance level 35%) — —.
Likely range 23%–99% (95% interval, resampling its 9 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
20 targets ranked
Held-out test: 61% of 51 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.16; matched chance level 40%) — —.
Likely range 19%–91% (95% interval, resampling its 10 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
8 targets ranked
Held-out test: 57% of 150 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.17; matched chance level 37%) — —.
Likely range 21%–88% (95% interval, resampling its 22 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
silver · district window · statewide
19 targets ranked
Held-out test: 40% of 250 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.08; matched chance level 28%) — —.
Likely range 21%–60% (95% interval, resampling its 36 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
rare earths · district window · statewide
6 targets ranked
Held-out test: — of 1 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — too few known producers to test.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
lithium · district window · statewide
12 targets ranked
Held-out test: — of 4 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — too few known producers to test.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
cobalt · district window · statewide
20 targets ranked
Held-out test: — of 11 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — too few known producers to test.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
copper · district window · statewide
8 targets ranked
Held-out test: — of — known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — too few known producers to test.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
To protect this area, its outline is not shown here; it is shared with the briefing.
gold · district window · statewide
19 targets ranked
Held-out test: — of 1 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — too few known producers to test.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
To protect this area, its outline is not shown here; it is shared with the briefing.
copper · district window · statewide
4 targets ranked
Held-out test: 38% of 2,427 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — not yet re-checkable from the files on disk.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
lithium · district window · statewide
4 targets ranked
Held-out test: 40% of 15 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — not yet re-checkable from the files on disk.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
molybdenum · district window · statewide
13 targets ranked
Held-out test: 44% of 119 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.32; matched chance level 40%) — —.
Likely range 28%–60% (95% interval, resampling its 66 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
cobalt · district window · statewide
17 targets ranked
Held-out test: 36% of 25 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — not yet re-checkable from the files on disk.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
silver · district window · statewide
57 targets ranked
Held-out test: 41% of 844 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.47; matched chance level 40%) — —.
Likely range 29%–53% (95% interval, resampling its 115 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
copper · district window · statewide
23 targets ranked
Held-out test: 41% of 1,309 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — not yet re-checkable from the files on disk.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
copper · district window · statewide
53 targets ranked
Held-out test: 45% of 728 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.23; matched chance level 40%) — —.
Likely range 28%–60% (95% interval, resampling its 95 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
17 targets ranked
Held-out test: 40% of 3,299 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — not yet re-checkable from the files on disk.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
copper · district window · statewide
11 targets ranked
Held-out test: 50% of 633 known producer records fell in our top-ranked 20% of ground.
Chance test: beats strict chance (p = 0.044; matched chance level 40%) — —.
Likely range 38%–61% (95% interval, resampling its 164 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
nickel · district window · statewide
7 targets ranked
Held-out test: 52% of 29 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.25; matched chance level 40%) — —.
Likely range 17%–77% (95% interval, resampling its 14 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
copper · district window · statewide
6 targets ranked
Held-out test: 40% of 1,242 known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — not yet re-checkable from the files on disk.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
51 targets ranked
Held-out test: 47% of 746 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.19; matched chance level 40%) — —.
Likely range 29%–63% (95% interval, resampling its 85 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
molybdenum · district window · statewide
51 targets ranked
Held-out test: 83% of 66 known producer records fell in our top-ranked 20% of ground.
Chance test: beats strict chance (p < 0.001; matched chance level 40%) — —.
Likely range 64%–97% (95% interval, resampling its 22 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
gold · district window · statewide
65 targets ranked
Held-out test: 48% of 104 known producer records fell in our top-ranked 20% of ground.
Chance test: not distinguishable from chance yet (p = 0.21; matched chance level 40%) — —.
Likely range 27%–67% (95% interval, resampling its 39 deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.
rare earths · district window · statewide
15 targets ranked
Held-out test: — of — known producer records fell in our top-ranked 20% of ground.
Chance test: not measurable (—; matched chance level —) — too few known producers to test.
Likely range —–— (95% interval, resampling its — deposit clusters).
Too few known producers here to test the ranking, so we make no accuracy claim.