The trust mechanism · in numbers
Data quality, measured
Our error rates, with their denominators — and one rule underneath: suspect data is flagged, never silently corrected, because the geologist who signs a model is accountable for every number in it.
Evidence base: pilot corpus — 2 companies · 13 drill-table PDFs · 30 hand-verified intervals (Mt Isa district, Jun–Jul 2026). Expanding Aug–Sep 2026 to 20+ companies and 300+ verified intervals; ground-truth fixtures and precommitment hashes publish with it.
Measured results
| What we measured | Result | Denominator / method |
|---|---|---|
| Full-corpus extraction run | 377 PDFs (two companies, 5 years) → 13 carried drill tables → 41 collars, 218 assay intervals | every announcement fetched — none hand-picked |
| Audit findings on that corpus | 73 problems → 0 errors, 1 warning | six-dimension audit: accuracy · completeness · timeliness · consistency · validity · uniqueness |
| The surviving warning | a real grade difference between two announcements of the same interval — left flagged for a geologist | flag-never-correct, working as designed |
| Cross-announcement duplication | 136 → 95 intervals after merge/dedup | the same holes republished across announcements; provenance retained |
| Import-safe identity | 84 duplicate-ID collisions → 0 | incl. one hole ID shared by two collars 17 km apart; delivery-unique IDs, source ID kept as an alias |
| Template extraction recall | 30/30 intervals · 95% CI [88%, 100%] (n=30) | hand-verified fixtures, regression-tested; measures the template on its own format family. Unseen layouts are loud-flagged, never silently parsed — cross-format performance is what the expansion measures |
| Machine-readability of the record | core intercepts + grades ~95%; full collar geometry ~50% per announcement | measured against a hand-verified corpus; geometry assembled across announcements |
| Legacy open-file reports | 0/18 matched our modern templates | pre-JORC layouts — which is why legacy extraction runs as a separate, human-verified tier, not a template product |
| Open-file ingest, deduplicated | one legacy project: 764 collars / 7.7 k assays; audit problems 893 → 7 | interval dedup on ingest |
| Gaps we publish, not hide | one prolific operator: 196 announcements, 0 collars in the showcase area — and that is correct | all 196 spatially checked: their drilling clusters 20–70 km north of the box. Holding a permit ≠ drilling it |
The same accounting, drawn
Documents — where the corpus narrows
One run · 2 companies · 5 years. Units: announcements (PDFs).
Most announcements carry no drill table at all — the narrowing is the record's shape, and every drop is accounted for, never silently skipped.
Intervals — from extracted to delivered
From those 13 PDFs. Units: assay intervals.
The delivered number is smaller than the extracted one — because the same hole announced three times must arrive as one.
Problems found → problems left
The boring correctness layer
Initial release → quarterly → investor deck: recognised as one hole via a natural-key identity model; intervals merged with provenance to every source document.
Both values survive. The conflict is flagged for geologist review — history is never silently replaced.
A curated, human-maintained alias register. It does not invent links; unresolved holders are flagged. Automatic inference is deliberately not trusted yet.
Every delivered value links to its source announcement and page. The trust columns — confidence, holder match, completeness regime, source page — ship inside the export, not only on screen.
A two-tier check separates "above range — verify (possible bonanza)" from "implausible — likely a unit or source artifact", calibrated to published district grades. A real high-grade hit is not flattened by the rule that catches a units error.
Two guarantees, kept apart
Every delivered dataset ships with the full audit run on that data — its flags, its completeness ledger, its provenance. You are never asked to trust a global average about your own ground.
The table above, at its current pilot scope. It grows by protocol: sampling fixed before extraction runs, confidence intervals on every rate, fixtures downloadable, hashes published before verification.
The review ledger
Extraction is assisted; acceptance is human. Every batch below was locked with a SHA-256 hash of the raw predictions before review — so neither the sample nor the results could be quietly adjusted afterwards. Corrections found are published, because a review that never finds anything isn't reviewing.
First reviewed batch lands here (Aug–Sep 2026), with its precommitment hash and correction count.
Have a say in what gets verified next. Suggestions feed our public verification queue — name a company or draw an area and mention "verify" in the notes. Need an area compiled for your own work, guaranteed and model-ready? That's an area request.
Take our data apart
A demo can be faked; a hostile review cannot. We invite sceptical geologists to take a delivered dataset apart against the source announcements. Findings are published here regardless of outcome — including the errors found.
First external review: being arranged (Aug–Sep 2026) — results will appear on this page.