04 — Evidence
The numbers speak without qualification
Concentration Power
Model concentrates known deposits into 0.17% of the search space at peak threshold.
Capture Rate
Known gold occurrences retained within the top 20% of model scores.
Holdout Accuracy
Classification accuracy on spatially blocked holdout sets never seen during training.
Statistical Significance
Permutation test against spatial null. Not luck, not autocorrelation.
Effort-Adjusted Wilcoxon
Model beats exploration bias with absurd statistical certainty.
Major Producers
Ontario’s major gold-producing camps correctly identified by the framework.
Ontario Shield | Independent holdout protocol.
Validated against known deposits.
01 — Insights
Insights
Occasional observations on methodology, markets, and the subsurface. We publish when we have something worth saying.
The Problem with Prospectivity Maps
Summary
Most prospectivity models train on known deposits and predict more of the same. They are detection-biased — they find ground that looks like ground that has already been found. The result is a map of exploration history, not geological potential.
Detail
Our framework separates the question of geological favorability from the question of observational coverage. A two-layer spatial model estimates where minerals are likely to exist (the truth layer) and where exploration has been sufficient to find them (the detection layer). The frontier — ground that is geologically favorable but historically under-examined — only becomes visible when you decouple these questions. Conventional prospectivity mapping cannot see it because it was never designed to look.
On the Misuse of ‘AI’ in Exploration
Summary
The mineral exploration industry has developed an enthusiasm for artificial intelligence that outpaces its understanding of what these tools actually do. The result is a growing gap between marketing language and operational reality.
Detail
Disciplined application looks different from the marketing version. It means automatic feature selection that identifies which of 48 geological variables actually carry information, rather than an operator choosing inputs based on intuition. It means Bayesian posterior estimation that quantifies uncertainty, rather than a binary prospective/non-prospective classification. It means explicit bias modelling that accounts for where people have looked, rather than treating detection history as ground truth. The model serves the geologist. Never the reverse.
Why We Focus on Ontario and the Canadian Shield
Summary
There is a persistent belief that the best remaining discoveries lie in frontier jurisdictions with minimal exploration history. This is sometimes true. It is more often a justification for avoiding the harder question: why hasn’t this ground been found in Ontario, where people have been looking for over a century?
Detail
Ontario and the Canadian Shield offer something that frontier regions cannot: data density. A century of systematic geological surveys by the OGS, decades of geochemical sampling, and comprehensive geophysical coverage create a substrate rich enough for computational methods to extract meaningful signal. The opportunity is not in going where no one has looked. It is in looking differently at what everyone has already seen across the Abitibi, Red Lake, and Sudbury.
Interested in how we can apply our process to your ground?
Confidential · By Invitation