Our Method
AI isn't our tool.
It's our founding geologist.
Fable is AI-first by construction, not conversion. The prospect models were built before we pegged our first ground.
01 / How We Explore
Four steps. In this order. Every time.
Read everything
Our models ingest the complete public geoscience estate: state-wide geochemical compilations running to tens of millions of assays, every open-file exploration report filed over sixty years, airborne magnetics, gravity, electromagnetics and radiometrics. They normalise a record collected by hundreds of companies across decades into one searchable, statistically honest dataset.
Screen without prejudice
Machine screening tests every square kilometre against locally-calibrated anomaly thresholds, not against what a human expects to find there. The models carry no district folklore. They rank ground purely on evidence, which is precisely how overlooked provinces get found.
Prove the machine before trusting it
Every commodity model must first re-discover the known fields blind, the historic districts every geologist can name. Only when a model recovers the proven endowment without being told where it is do we let it rank the unknown. Discovery claims are cheap; control recovery is our internal standard of proof.
Then send geologists
AI proposes; geology disposes. Every machine-generated target passes a human verification ladder of source-data audits, historical drill-record reads, regolith and cover screening, structural interpretation and field checks before a dollar of ground spend. Speed comes from the machines; conviction comes from the rocks.
02 / The Asymmetry
A conventional junior evaluates one district per field season. Our stack evaluates a state in weeks and re-runs it overnight as new data lands. In a first-in-time tenure system, that speed is the entire competitive advantage, not a convenience. We apply for ground others haven't yet noticed is open, over anomalies others haven't yet noticed exist.
03 / The Kill Tests
Most of the work goes into killing our own targets.
A screen that only ever says yes is not an instrument, it is an advertisement. So the second half of the method exists to destroy the output of the first half, and every test is written down before the answer is known.
Freeze the instrument before testing it
Each commodity model is fixed and recorded before it is run against the validation set, so a score cannot be quietly tuned afterwards to flatter the result. A model adjusted once it has seen the answer has not been tested, it has been fitted.
Run the failures, not just the finds
Recovering known deposits proves very little on its own, because a generous threshold recovers everything. Every model is also run over ground that was properly explored and found barren. A model that flags the discoveries and the duds at the same rate has told us nothing, and it is recorded that way.
Go back to the original page
Any historic number a target leans on is read at its source, including the appendix sheets and microfiche behind the typed summary tables. Units, sample type and depth are checked before the value is believed. Promising intercepts have turned out to be a units error, a different hole, or a figure caption misread for a result.
Ask why the last holder walked away
Ground that was explored and dropped carries a stated reason in the open file, and we go and find it. Where a previous operator abandoned a position on the same grounds our target depends on, that makes the target weaker, not stronger, and it is weighted down accordingly.
Test the boring explanation first
Many anomalies are a surface process rather than a deposit: a drainage train, a regolith blanket, a landform, a survey artefact. The dull explanation is tested and excluded before the interesting one is allowed to stand.
Price the test, not the prize
A target only survives if there is a cheap and decisive way to prove it wrong. Where the only way to find out is an expensive drill program, it waits behind the ones that can be settled with an assay or a traverse, whatever it scored.
Most targets do not survive this, and neither have several of our own methods. Screens we built, registered and then had to retire stay on the internal record with the reason they failed, because a method that has never produced a negative result has never been tested either. Killing a target at the desk costs a few days of reading. Drilling one that should have been killed costs a season and a budget.
04 / Where It Runs
The same engine drives all three subsidiaries, calibrated per commodity system and proven on that system's known fields before it ranks anything new.