# When AI Starts Writing the Reserve Statement, Watch the Footnotes
A Toronto outfit called Stratum AI now claims its deep-learning models underpin the publicly reported mineral resource estimate at an operating gold mine — St Barbara’s Simberi operation in Papua New Guinea. The company calls it one of the first cases of its kind in mining. Maybe. But before anyone starts rerating junior explorers on the promise of “AI-defined resources,” here’s what this really means — and what it doesn’t.
## The facts, stripped of the press-release varnish
Stratum spent four-plus years with St Barbara at Simberi building models that classify material type — oxide, transitional, sulphide — and weathering profiles. That’s the actual job. Not sci-fi geology, just pattern recognition on data mines already generate: drill logs, assays, grade-control samples.
The practical payoff came first on the operational side. Simberi runs a carbon-in-leach plant built for oxide ore. Feed it the wrong material and you’re burning mill liners and producing waste. Stratum’s models told the mine which ore could go through the CIL circuit and which should be stockpiled for future sulphide treatment. Less waste, less wear, lower cost. Fine. That’s a legitimate use of machine learning, and frankly the kind of unglamorous application that actually earns its keep.
After the models were validated against grade-control reconciliation — the mining equivalent of marking your homework against actual production — St Barbara made them its preferred tool for material-type classification. They fed into the 2024 resource and reserve update and the 2025 Feasibility Study.
The headline numbers: 143,000 ounces of previously overlooked treatable material identified in May 2024, extending oxide processing by 17 months. A 40% bump in proved and probable reserves announced a month later. The AI-supported resource estimate is now reported under the JORC Code — the standard baked into ASX listing rules.
St Barbara released the study on December 10, 2025, the same day it announced a deal with Lingbao Gold Group valuing the New Simberi expansion at roughly A$800 million, or about US$562 million.
## This isn’t the first time we’ve seen this circus
I’ve watched technology narratives arrive in mining before. Heap leach in the 80s. Biox in the 90s. “Big data” in the 2010s. Every wave comes wrapped in the promise of unlocked ounces and repriced assets, and every wave delivers something real — just never as fast, broadly, or cheaply as the pitch deck suggests.
So when a CEO says “we see this as a significant milestone in the adoption of AI in mining,” my default is to reach for the reconciliation reports, not the press release.
Credit where due: the sequence here is disciplined. Stratum piloted the models, tested them against operational data, and — this is the part that matters — validated against reconciliation before anyone put the numbers in a JORC-compliant public document. That’s how you earn trust with a model. Most of what passes for “AI in mining” never gets that far. It stays in the slide deck.
But keep your eyes open. A mineral resource estimate is a probabilistic statement, and the model is only as good as its training data. If Simberi’s geology shifts where the drill holes are sparse, the model extrapolates — same as a human geologist, just with more confidence and less accountability. JORC compliance means a Competent Person signed off. That’s a real backstop, but a signature isn’t a guarantee. Regulators don’t audit algorithms; they audit whether a responsible person believes them.
## The numbers deserve scrutiny, not cynicism
A 40% reserve increase is a big move. The skeptical question isn’t whether the ounces exist — it’s whether they were always there and better classification finally found them, or whether the model moved the boundary of what “counts.” Both are legitimate outcomes. The first is genuine value creation from better information. The second is a definitional shift dressed up as discovery. The distinction matters when the same resource estimate is anchoring an A$800 million transaction valuation.
The 143,000 ounces and 17-month mine-life extension, similarly, are real only to the extent the metallurgy behaves. Classifying material as treatable in the CIL circuit is a bet on recovery rates holding. Reconciliation over the next few quarters will tell the truth, as it always does.
## Where the real opportunity sits
The genuinely interesting point buried in this story isn’t AI — it’s data. Mining companies sit on decades of drill core, assays, and production records that are largely gathering digital dust. Any technology that converts that sunk cost into better classification, better blending decisions, and less mill damage is worth real money. It doesn’t require believing in an AI revolution. It requires believing that better decisions at the ore-sorter and the mill feed beat another decade of expensive drilling.
For investors, the takeaway is unglamorous: watch the reconciliation. If St Barbara’s production keeps matching the AI-informed estimate, this methodology spreads, and the companies quietly building validated geological models — not the ones slapping “AI” on their investor pages — will win the follow-on contracts. If the model drifts from reality, this becomes a case study in why regulators move slowly.
One more thing worth flagging: the timing. A JORC-compliant AI-backed resource, a feasibility study, and an A$800 million strategic deal all landing the same day is excellent narrative management. It doesn’t make anything wrong. But seasoned investors know that the best stories tend to arrive pre-packaged with the transaction they’re meant to support.
The bottom line is simple. Better ore classification is real value; the label “AI” is not. Judge the ounces, the metallurgy, and the reconciliation — nothing else.
