Exploration targets you
can actually drill.
GeoCore runs its validated prospectivity engine on your data and delivers only the targets that survive honest spatial testing. No inflated AUCs. No artifacts dressed up as discoveries.
Built for scrutiny. Spatial validation · confound correction · NI 43-101 / JORC-aware reporting
An 0.82 that means nothing.
Prospectivity models routinely report ROC-AUC of 0.8–0.9 and produce confident target maps that move millions in drilling budget. That performance is often an artifact.
Geoscience data is dominated by smooth spatial fields: terrain, regional gravity, magnetics. A model can score high simply by memorizing where deposits sit inside a region, then fail completely on ground it has never seen. Ordinary random-split validation hides the failure entirely.
The result is that drill budgets follow map patterns, not ore. Exploration teams need targets that have survived honest spatial testing, not another model that scores well in the lab and fails in the field.
The Dominance Diagnostic Engine.
GeoCore doesn't just rank targets, it audits its own model. Before anything ships, the DDE scores every evidence family on three axes that separate real geology from geometry.
Who's holding the model
Share of model importance held by an evidence family, magnetics, gravity and their derivatives grouped as one GEOPHYSICS family, so a dominant proxy can't hide by splitting itself across correlated columns.
What breaks if it's gone
Out-of-sample PR-AUC lost when the whole family is ablated and the model retrained under spatial block cross-validation. High dominance with near-zero necessity means the signal is a redundant proxy, not load-bearing.
Does it hold across ground
Consistency of a family's importance share across contiguous spatial folds. A family that only matters in one fold is describing that fold, not the deposit.
No feature is suppressed by fiat. SUPPORTED DRIVER and SUBTLE BUT ESSENTIAL families are never touched, dominance earns its place when it's necessary, not because it's large.
Validated against ground it never saw.
Held to the same standard we hold every model to: contiguous spatial folds, a fully hidden quadrant, and deposits the engine never touched during training.
Spatial block CV ROC-AUC
Strict contiguous geographic folding, no random splits.
Hidden deposits
Scored blind in a completely unseen geographic quadrant.
Held-out hit rate @ ≥0.50
Recovery rate against those 49 hidden deposits.
Terrain-matched contrast AUC
Signal separation from background terrain, not topographic memorization.
Verdict: VALIDATED, held-out deposits recovered at decision grade. This is the same class of dataset shown above scoring ARTIFACT before remediation: the DDE closed loop is what changed the outcome, on real deposits, not a cherry-picked run.
Rigor, made operational.
Five stages GeoCore runs on your data so every target we deliver has survived the tests the industry skips.
Geolocate
Every sample deposit and background keeps its true location. No random splits, no spatial leakage into the result.
Hold out
GeoCore scores prospectivity against a whole region the engine never trained on. The only honest test of new ground.
Attribute
Trace apparent signal to its physical source: terrain roughness, regional gravity, sampling bias. If it's an artifact, we find it.
Correct
Terrain-matched sampling plus the Dominance Diagnostic Engine ablate any evidence family that dominates the model without earning it, then retrain before a target ships.
Deliver
Targets come with calibrated confidence and depth-to-source. If nothing survives, we say so plainly, no dressed-up artifacts.
How GeoCore filters real targets.
A real run on public data. This is how GeoCore separates drill-worthy targets from spatial artifacts before your rig moves.
By every in-region measure, this is a fundable model. Most tools would ship this map and call it done.
real deposits sit in low-probability ground. the targets track the field, not the ore.
These targets would have moved drill budget. GeoCore filtered them out before the rig moved, that's what honest spatial testing delivers.
Why GeoCore targets hold up.
Spatial truth
Block cross-validation and out-of-region holdout baked into the engine, not optional, not quoted as a footnote. Every target survives the test others skip.
Artifact-free delivery
Terrain and regional-field confounds are detected and stripped internally. What you receive are targets driven by geology, not by smooth geophysical fields.
Defensible by design
Calibrated uncertainty, full provenance, and reporting built for NI 43-101 / JORC scrutiny. Targets you can stand behind at the board table.
Targets your drills can follow.
We're onboarding a small group of design partners, exploration teams and investors ready for targets that have survived honest spatial testing. Bring your data. We run the engine and deliver only what's real.
Request a pilot ↗