Digitizing Soviet geological archives means converting typewritten otchet reports, GKZ reserve protocols, 1:200,000 and 1:50,000 map series, field books and assay ledgers into structured, coordinate-correct data. That requires Cyrillic OCR tuned for technical text, Pulkovo 1942 to WGS-84 reprojection, and GKZ category mapping to JORC, NI 43-101, KAZRC, SAMREC, PERC 2021 or S-K 1300, so decades-old fieldwork becomes usable again.
A typical archive of a former Soviet republic holds five recurring document families: bound otchet (report) volumes summarizing a field season or exploration stage, GKZ protocols recording the reserve approval decision of the State Commission on Reserves, topographic and geological map series at 1:200,000 and 1:50,000 scale, handwritten field books kept by the geologist on site, and laboratory assay ledgers listing sample results by number. These were produced continuously from the 1930s through the early 1990s, and in most republics the volume runs into the millions of individual sheets, bound folios, map plates, magnetic tapes and index cards.
| Document type | What it contains | Typical physical form |
|---|---|---|
| Otchet (report) volumes | Narrative exploration or prospecting report: geology, methods, results, reserve calculation | Typewritten bound folio, often 100 to 400 pages, with fold-out map appendices |
| GKZ protocols | Formal reserve approval decision, category (A/B/C1/C2/P) assigned per deposit block | Typed protocol sheets with commission signatures and stamps |
| 1:200,000 and 1:50,000 map series | Regional and detailed geological, topographic and mineral occurrence maps | Large-format printed or hand-colored sheets, Gauss-Kruger grid |
| Field books | Handwritten outcrop descriptions, sample numbers, sketch sections | Pocket-sized notebooks, often faded pencil or carbon-copy ink |
| Laboratory assay ledgers | Sample number, method, grade result, cross-referenced to drill or trench interval | Ruled ledger books or carbon-copy assay certificates |
None of this was built to be machine-readable. It was built to satisfy a state reserve-approval process, and it still carries the internal logic of that process: category codes instead of confidence intervals, local coordinate grids instead of geographic ones, and narrative prose instead of structured fields.
Standard document-OCR pipelines fail on this material for five specific reasons, not one.
A working pipeline runs six stages in sequence, and skipping any one of them produces a dataset that looks digitized but is not actually usable for exploration decisions.
A single misread digit in a coordinate or an assay grade does not fail loudly, it fails quietly, as a drill target placed a zone-width away from where the historical sample actually was. Cross-checking against modern data, and flagging what cannot be cross-checked, is what makes the output usable for a drilling decision rather than just searchable.
As of May 2026, 4,859 documents had gone through this process, and the volume itself changed how the pipeline works, not just how much data it holds. Recurring OCR failure modes in specific decades and typewriter models became identifiable and correctable in bulk rather than case by case. Abbreviation and unit conventions that shifted between the 1950s, 1970s and 1980s stopped being one-off exceptions and became a lookup table applied automatically by document date. Coordinate-zone determination, initially a manual check per map sheet, became reliable enough to automate for the majority of sheets, with manual review reserved for edge cases near zone boundaries or damaged grid labels. That structured, cross-checked corpus is also the base layer behind a 76% AI-ranked drill hit rate on ranked targets, since the ranking model has historical assay and lithology data to compare a new target against instead of only current-season drilling.
The direct payoff of getting this right is greenfield-cost discovery on ground that already has decades of fieldwork sitting in an archive. A license block with an existing otchet report and GKZ protocol already has drill intervals, assay results and a geologist's structural interpretation on file; the cost of a new discovery there is the cost of digitizing and re-interpreting that record, not the cost of a first-pass regional exploration program from bare ground. For jurisdictions where the Soviet-era exploration effort was intensive, that difference is the entire economic case for revisiting old licenses instead of pursuing untested greenfield ground. It only works, though, if the historical data is actually structured and coordinate-correct rather than a pile of scanned images that still requires a geologist to read page by page.
Three groups run into this problem for different reasons.
Across all three, the practical guide to the category conversion itself is covered in the GKZ to JORC conversion walkthrough, and general CRIRSCO code comparisons are in the methodology reference.
CoreElement AI, founded in 2024 by Daniel Tonkopiy and Zhambyl Suraganov, runs a Soviet GKZ A/B/C1/C2/P conversion module as one of 22 modules across its exploration platform, combining Cyrillic-tuned OCR, Pulkovo 1942 to WGS-84 reprojection and GKZ category mapping with the CRIRSCO-aligned codes: JORC 2012, NI 43-101, KAZRC, SAMREC, PERC 2021 and S-K 1300. The module has processed 4,859 digitized documents as of May 2026, runs two pilot deployments with large enterprise mining operators in Kazakhstan, and is built with advisors from Stanford, MIT and Cornell Tech alongside academic collaboration with UC Berkeley and Nazarbayev University. It is an NVIDIA Inception member, a Google for Startups Cloud participant, and an Astana Hub resident.