Afrimintel · Platform · Decision Views · Countries · EITI · Insights · Methodology · Audit Log
Insights · 12 June 2026 · Editorial responsibility: Nikesh Patel
We built a rule to infer which country a mine sits in. It put a South African mine in Botswana — twice, two different ways. This is why we killed it, and why 70 of our assets now say Absent where a country name could go.
Our country pages needed asset tables. Every mining analyst knows Jwaneng is in Botswana and Kamoa-Kakula is in the DRC. The obvious build: scan each asset record's text for country names and attribute accordingly. Thirty-nine of our records state their country in the record name itself — those attribute with zero inference. For the rest, a single-country mention across the record's fields looked like a safe rule.
Venetia is De Beers' diamond mine in Limpopo, South Africa. Our record's text never says "South Africa" — but its operator line names the Government of the Republic of Botswana, because Botswana genuinely holds 15% of De Beers Group. One country mentioned, zero mentions of the host. The rule attributed Venetia to Botswana with full confidence.
We tightened it: only a sovereign equity holder in the operator field counts — surely a government holding a stake is the host state. Kamoa-Kakula's "DRC Gov (20% non-dilutable)" attributes correctly under that rule. And then the same Venetia record disproves the premise: Botswana's stake in De Beers Group is sovereign equity in a company operating someone else's mine. Governments hold equity across borders. The rule only dodged Venetia by a wording accident — "Republic of Botswana" instead of "Botswana" — and a rule that survives by luck is already dead.
Name-stated attribution only. Thirty-nine assets appear on their country pages because their records say where they are. Seventy say Absent — including Jwaneng, which every human reader knows is in Botswana. The record doesn't say it, so the page doesn't either. The fix is unglamorous: a sourced country field added to each silent record, one verification at a time. The full forensics, including both failed rules, are on the public audit log under v2.11.0 — recorded before the fix was declared complete, because that's the order the discipline requires.
Text inference fails in exactly the way that hurts most: confidently, on records that mention a foreign country once and the host country never. Any data product doing entity-attribution at scale has Venetias in it — assets filed under their financier's jurisdiction, their offtaker's, their parent's listing venue. Most platforms will never know which ones, because the inference that created the error leaves no trace. A three-state discipline does the opposite: it converts every uncertain attribution into a visible Absent, which is a work queue, not a defect. The discount on African mining data is built from a thousand silent Venetias. We'd rather show you the seventy.
Sources: Afrimintel audit log v2.11.0 (the full inference-tier forensics); De Beers Group ownership structure (Anglo American 85% / Government of the Republic of Botswana 15%, public record); Afrimintel country pages at /country/. Every claim in this article about our own system is verifiable on the audit log.