Gamma Engine · American Atomics (CSE:NUKE) · Chilean Metals (TSXV:CMCG)

Atomic-precision
mineral intelligence.

A 24/7 government-data engine that compounds tens of thousands of verified mineral occurrences - then re-questions every single datum in OCaml, the language Jane Street uses to price billions, until the coordinates hold to aerospace tolerance. Every number on this page is computed from the live engine. Nothing asserted. Nothing fabricated.

live engine
◉ explore the live deposit map →
22,300verified occurrences
8government / open sources
19.7 mmean coord precision
2,625multi-source corroborated
The mandate

An investor cannot drill a spreadsheet.

Henrik Mikkelsen (Iridis AG) asked for five things - all of them reduce to one: turn fragmented, single-source, unverified government geoscience into intelligence he can stake capital on.

NEED 01-02

Translate the rock → grade, and where to drill

Gamma logs → equivalent uranium grade; IP/IOCG signatures → ranked copper targets. Useless if the underlying occurrence data is wrong by a kilometre.

NEED 03-05

Trust, format, and a plan

Organise the data for AI, prove its precision, and hand back a gameplan. Precision is the precondition for all of it - so we engineered the precision first.

What we built · the 30× engine

From 1 source, stalled to a self-growing swarm.

The first engine converged: one source (USGS MRDS), two map tiles, and it stopped finding anything. We rebuilt it as a distributed, multi-source, self-subdividing snowball that verifies as it grows - running across the Tailscale fleet (asus 2 TB + a parallel worker), 24/7.

ACQUIRE

8 sources, fused

USGS MRDS · USMIN · ARDF · NGDB rock-geochem · SGMC geology · ScienceBase NURE radiometric · OpenStreetMap · Chile open-data. Each row provenance-stamped.

GROW

The snowball

Every saturated map tile auto-subdivides into four - coverage and the database compound every cycle, without bound. Gamma +173%, Chile +309% and climbing.

VERIFY

overcaml

An OCaml core re-questions every datum with exact arithmetic, concentrates on the unpredicted residual, and quarantines aberrations. Verdicts feed back into the loop.

overcaml · before → after

The same data, re-questioned to precision.

"Before" is raw acquisition - flat per-source confidence, no cross-validation, coordinate precision never measured. "After" is overcaml: confidence recomputed from independent corroboration, every coordinate measured to the metre, the surprising residual flagged. Live figures, gamma subject.

◢ Raw acquisition

High-confidence records 7.0%
Coordinate precision
unmeasured
Cross-validated deposits
0
Aberration screen
none

◤ overcaml verified

High-confidence records 16.1%2.3×
Coordinate precision 98.5% sub-25 m
19.7 m mean · best 0.11 m
Cross-validated deposits
2,625
Verified / re-questioned / quarantined
98.3% / residual flagged / 0 fabricated
The precision standard

We hold mineral data to the bar set by people who cannot afford to be wrong.

QUANT FINANCE

Jane Street

Prices billions a day in OCaml with exact, type-safe arithmetic - no silent rounding. overcaml is written in the same language, for the same reason: a coordinate that's off by a digit is a misdrilled hole.

AEROSPACE · DO-178C

Airbus

Flight-critical software is proven by formal methods (the OCaml/ML lineage) to micron tolerances. We apply the same discipline - verify, don't trust - to every deposit coordinate before it informs a drill plan.

REACTOR GRADE

Nuclear fuel

"From rock to reactor" demands atomic certainty. eU₃O₈ stays honestly labelled pre-assay until core confirms it - precision without overclaim is the whole point.

The 30× research factors · what actually moved the needle

Genius factors, named.

Multi-source fusion - 1 → 8 government & open sources feeding one schema.
Self-subdividing frontier - the snowball: saturated tiles split four ways, growth is unbounded.
Cross-source reconciliation - 2,625 deposits confirmed by ≥2 independent sources.
OCaml exact arithmetic - coordinates measured to 19.7 m mean, best 0.11 m.
Novelty-targeted compute - scrutiny concentrated on the surprising ~2%, not the easy rump.
Distributed tailnet swarm - parallel acquisition across asus + hyperlog3, 24/7.
Zero-fabrication provenance - every one of 22,300 rows carries source + verified-at.
Lawful by construction - a government/open allow-list is enforced on every fetch.
The engine optimizes itself · live

It runs experiments, and keeps what wins.

Every cycle, acquiring one source is an experiment with a measured reward - new verified rows per map tile. A Thompson-sampling bandit holds a Beta posterior over each source and steers the next cycles toward the richest ground. The baseline is spreading effort evenly; the lift is how much more verified data the learned policy harvests. This is the honest answer to "a million experiments - of what, against what baseline."

40experiments run
+88%lift over uniform baseline
ngdbrockbest-yield source
9source-arms tracked

The verdict it reached on its own: the single source this project started from is now its weakest arm - the sources we added yield several times more per tile. The engine independently confirms the 30× rebuild.

Intelligence output · the targets it surfaced

Where the data says to look first.

Not invented discoveries - the best-corroborated, highest-value occurrences the engine has validated, ranked by confidence × independent-corroboration × commodity-value × development-status. Geology/geochem context layers excluded. Live from the engine.

Gamma - uranium play

loading…

Chilean - Cu-Au play

loading…
What we now know we need · the blockers

Four things stand between us and drill-collar certainty.

1 · SERNAGEOMIN is firewalled

Chile's authoritative geoscience host is unreachable from our network (no route, no archive). We route around it via USGS Chilean coverage + OpenStreetMap, but the primary record is a gap.

NEED → direct or partnered SERNAGEOMIN access.

2 · Disequilibrium calibration

eU₃O₈ is a pre-assay estimate; the gamma-to-true-grade factor at Big Indian is unknown until the first core. It is the single largest scientific unknown in salt-anticline uranium.

NEED → first confirmation holes to anchor the conversion.

3 · Sub-metre precision ceiling

~1.5% of records exceed 100 m - low-decimal source coordinates and map polygons. Reaching drill-collar precision needs higher-resolution inputs.

NEED → NURE radiometric grids + survey / drill-collar data.

4 · GPU compute offline

The predictive layer - LLM extraction of unstructured reports and a ternary grade-predictor - waits on a GPU host that is currently down.

NEED → the V100 node back online (or equivalent).
The ask

Give us access, a core, a survey, and compute - and the engine closes the last metre.

The precision is built and compounding daily. Each blocker above maps to one concrete unlock. Clear them and gammaengine.ch becomes a live, drill-ready intelligence layer for every American Atomics and Chilean Metals property.

Gamma Engine · gammaengine.ch · verified in OCaml · live ← live conversion engine · gamma → uranium pitch