How We Think
Discovery is not serendipity. It is the systematic reduction of geological uncertainty through evidence, iteration, and disciplined inference. The mineral exploration industry has operated for decades on a model that privileges narrative — geological stories told by experienced practitioners who synthesize maps, field observations, and intuition into targets. That model produced the deposits we know about today. It also produced a landscape where the vast majority of prospective ground has never been seriously examined, because human attention is finite and geological storytelling does not scale. Our methodology is built on five principles that govern every model, every score, and every capital allocation decision we make — not because we believe technology replaces geological expertise, but because we believe the next generation of discoveries will come from ground that narrative-driven exploration was never going to reach.
48.4°N · 81.3°W — Ontario Shield
Measurement Over Intuition
We do not interpret maps. We measure them. Our system evaluates a broad set of candidate geological variables and automatically determines which carry genuine predictive signal and which do not. No feature enters a score because someone believed it should, and no feature is excluded because it contradicts a prevailing geological narrative. This is a deliberate inversion of conventional practice, where an experienced geologist selects the variables they believe are relevant and builds a model around that selection. Variable selection is one of the highest-leverage decisions in any predictive system, and we remove it from human hands entirely — not because human judgment is worthless, but because human judgment cannot be applied uniformly across a large and heterogeneous geological province without introducing systematic bias toward the geological environments the operator has personally experienced.
Separation of Truth from Observation
Every mineral deposit dataset is contaminated by a fundamental confound: it records where deposits have been found, not where they exist. The known mineral occurrence record for any jurisdiction reflects a century of exploration decisions as much as it reflects underlying geology. Areas near roads, near towns, near existing mines, and near historical survey corridors have been examined intensively. Areas that are remote, covered by thick drift, or simply unfashionable have not. Conventional prospectivity models train on this biased record and inevitably learn to predict where people have looked rather than where gold is present. Our architecture addresses this directly by separately estimating geological prospectivity — the probability that mineralization is present — and observational coverage — the probability that existing mineralization would have been detected given historical exploration patterns. This separation is imposed by the mathematical structure of the model, not by the operator, which means it cannot be overridden, tuned away, or selectively applied. The result is a system that answers the question we actually care about: not “where have people found gold?” but “where does the geology suggest gold should be, regardless of whether anyone has ever looked?”
Calibrated Uncertainty, Not Point Estimates
Exploration is inherently uncertain, and any system that produces rankings without quantifying the confidence behind them is gambling blind. Our framework produces posterior probability distributions for every cell in the grid — not point estimates, not binary classifications, not confidence-free rankings that say one target is “better” than another without saying whether either one is worth staking. Every probability we generate has a credible interval behind it, and every target carries an explicit uncertainty width that tells us how much conviction could shift if the data were slightly different. This matters because exploration capital is finite and ours is no exception. A target with a high expected score but wide uncertainty requires a different position size than a target with a moderate score and narrow uncertainty. By quantifying this distinction rather than hiding it, we allocate capital rationally — concentrating where conviction is high and diversifying where it is not.
Uniformity at Scale
A method that works on one greenstone belt is an anecdote. A method that works uniformly across an entire geological province is a system. Our model operates identically across the full extent of its target jurisdiction, applying the same mathematical framework, the same selection process, the same convergence criteria, and the same validation protocol to every cell in the grid. Sub-provincial variation is accounted for within the model structure, but the inference machinery itself is uniform. This is not a design convenience; it is a methodological requirement. Any system that is tuned, recalibrated, or manually adjusted on a belt-by-belt basis introduces operator degrees of freedom that compromise reproducibility and make it impossible to compare conviction across regions. A score in one district means the same thing as a score in any other. Scale demands discipline, discipline demands full automation, and automation demands that every step from data ingestion through scoring and capital deployment is executed by code that runs identically every time without manual intervention.
Validated, Not Claimed
Every model output passes through a multi-stream validation protocol before it informs a staking or acquisition decision. This is not a single train-test split or a cross-validation score reported in a footnote. It is a structured sequence of progressively harder statistical tests designed to answer the specific questions that matter for deploying exploration capital. Can the model generalize to regions it has never seen? Can its performance be explained by spatial autocorrelation alone? And hardest of all — within areas of comparable exploration maturity, does the geological signal still discriminate, or is the model simply recapitulating the pattern of historical activity? We designed the validation framework around the failure modes we are most concerned about — spatial leakage, effort confounding, and overfitting to the training distribution — because deploying capital on a model that passes easy tests and fails hard ones is worse than not deploying at all. We do not act on scores we cannot defend under adversarial scrutiny.
From Raw Signal to Conviction
Data Assembly
Multi-domain geoscientific datasets are assembled from public geological survey archives, government data releases, and open-source platforms. Each data domain passes through a dedicated extraction pipeline that handles coordinate harmonization, spatial resampling, missing-data encoding, and quality verification. All features are harmonized to a uniform spatial grid, ensuring that every cell in the province is described consistently. Every input is catalogued, version-locked, and integrity-verified before it enters the model.
Quality Firewall
Before any data source enters a production model, it must pass a fixed set of quantitative criteria that test for statistical significance, access bias, geographic stability, and contamination by historical exploration patterns. Data sources that correlate with where people have drilled — rather than where gold exists — are automatically blocked. This firewall is non-negotiable: it runs identically on every data source, the criteria are fixed before the data are examined, and no override mechanism exists. It is the single most important quality-control gate in the system, because the most dangerous inputs are the ones that look most useful.
Inference
The core model separately estimates geological prospectivity and observational coverage, then combines them into a single calibrated score for every cell in the grid. The model automatically identifies which variables carry genuine signal and suppresses those that do not. Convergence is verified through standard Bayesian diagnostics. The output is a full posterior distribution — not a point estimate — for every cell, capturing both our best estimate and our uncertainty.
Attribution
Every cell score is exactly decomposable into the geological factors that drive it. This is not an approximation or a post-hoc explanation — it is a mathematical property of the model architecture. Attribution tells us not just where to deploy capital, but why the model holds conviction there, enabling geological validation of every position we take. It also differentiates between targets — two cells with identical scores may be driven by entirely different geological signatures, implying different hypotheses and different follow-up programmes. Attribution is what turns a ranked list into an exploration strategy.
Frontier Targeting and Capital Deployment
Cells are scored and clustered into spatially contiguous target zones, tiered by model conviction. The highest-value positions are frontier targets: ground where the model indicates strong geological favourability but minimal historical examination. These are cells that score highly not because someone has already found something nearby, but because the geological signature matches patterns associated with known deposits in entirely different parts of the province. We acquire ground where our conviction is highest and the market’s attention is lowest. Every position is backed by a complete internal dossier — score, uncertainty interval, attribution profile, exploration-history context — and the entire pipeline from source data to staking decision is version-locked, integrity-verified, and fully reproducible.
By the Numbers
0+
Square kilometres modelled
0
Integrated datasets
0
Frontier cells identified
0%
Deposit capture rate
We do not predict. We prepare. Every model is a hypothesis. Every target, a test.