Version DR-0.1 · published 31 August 2026 · effective from 31 August 2026 · supersedes nothing · machine-readable record: divergences.json
What this is. A per-asset record of where the public record disagrees with itself. Each record carries two Sourced values, the documents they come from, the basis and date of each, and whether a reconciling disclosure was located.
What it never does. It never states which figure is correct. It never says a party is wrong, misleading, or non-compliant. It publishes no score, ranking, league table or severity measure, and no comparative count of divergences per operator — that would be a ranking with extra steps. The reader draws the conclusion; this record supplies the ingredients. That refusal is the point, and it is what makes the output usable in a committee paper.
The platform already runs cross-feed consistency machinery against its own data. This layer is the same discipline pointed at the public record, with a disposition workflow and a publication bar.
Most divergences are ordinary and explain themselves: a newer study superseded an older one, or two figures are measured on different bases. Those are published too, and the reason is not padding. A surface that shows only problems is not a measurement, and a professional reader discounts it accordingly. The rare case — two current figures, same basis, no reconciling note anywhere in the documents searched — means something only because the ordinary cases are shown beside it.
| Class | Test | A finding? |
|---|---|---|
| superseded | A newer document explicitly replaces the older figure, with a dated restatement note. | No — this is disclosure working as intended. |
| basis difference | Same asset, different measure: reserve vs resource, 100% vs attributable, wet vs dry, pre- vs post-tax, differing reporting codes. | No. Most apparent contradictions die here. |
| vintage lag | Both figures were correct at their own dates; one surface has not caught up. | No. Where the lagging surface is ours, it is marked and corrected under the published SLA. |
| record vs consensus | The primary instrument states one figure; widely-reported secondary sources carry another. | Qualified — a finding about the citation chain, never about the issuer. |
| drift outside a study | A figure cited outside any completed study diverges from the last study figure, with no interim document. | Qualified — published as “no completed study carries this figure”, never as an accusation. |
| unreconciled | Two current figures, both Sourced to the claimant’s own documents, same basis and scope, no supersession note and no basis explanation located. | Yes. Expected to be rare. |
Rule of construction. A record may be classified unreconciled only after each of the five ordinary classes has been positively ruled out, with that exclusion recorded and the searched document set named. Absence of an explanation is not evidence of absence.
Before any two values are compared, the record must confirm from the documents themselves that they describe the same asset: operator, licence or tenement reference, and location. Where any of the three cannot be confirmed, the record closes and is retained as a negative record.
This gate exists because we have made this error. An internal flag once reported a single-asset capital-cost contradiction between US$712m and US$1,061m. The two figures belong to different assets in the same country — Koné (Montage Gold) and Assafou (Endeavour Mining), different operators, different deposits, different studies. Both records were correct, and the obvious fix would have overwritten a correct figure in the name of consistency. Same country, same commodity, similar magnitudes, adjacent in an alphabetical list: that is the shape of the trap, and a near-identical number is grounds for more suspicion, not less.
Every published record renders to the same pattern: Document A (dated) carries X on basis P. Document B (dated) carries Y on basis Q. A reconciling note was located at Z — or no reconciling note was located in the documents searched on this date, which are listed.
v0.1 is one asset, searched across five fields. Kabanga: reserve/resource, resource M+I, FID timing, development stage, operating status. Six in-scope fields are not yet searched — capex, NPV, IRR, ownership and control, government take, production guidance. The remaining five intelligence-grade assets are queued — Kamoa-Kakula, Manono, Balama, Colluli and Ngualla — and each is published only when its primary-document pass is complete.
Coverage is declared, never implied. “No divergence located” means none located in the fields and documents listed on the record, on the date stated. It is not a statement that none exists, and it says nothing about a field that has not been searched. This is the same discipline as the platform’s absence basis, which distinguishes a gap in the public record from a gap in our own verification: an unchecked field is not a clean field, and a layer that blurred the two would be committing the error it exists to surface. The searched and unsearched field lists ship in the feed under coverage.
Every African jurisdiction in the platform’s coverage, without exclusion — including the Democratic Republic of the Congo, Rwanda and the wider Great Lakes region. An earlier version of this layer excluded those jurisdictions. That exclusion was withdrawn by editorial ruling on 31 August 2026 and the withdrawal is recorded on the public audit log rather than applied silently. The reasoning is that this layer reaches no conclusion: a record states what two documents say, on what basis, as at what date, and whether a reconciling disclosure was located. It expresses no opinion and takes no position on any dispute. The platform’s standing constraint against editorial opinion and against commercial outreach in those jurisdictions is unchanged and unaffected by this ruling.
Afrimintel’s own divergences are published in this layer at the same bar, and are marked as ours. Of the five records in the v0.1 seed, three are ours: two vintage lags corrected on publication, and one unresolved resource-basis conflict left open with a resolution date, because it cannot be settled without reading the technical reports side by side and this record will not guess.
A layer that examines the public record must examine its own first, and be seen to. A contradiction layer that has never published its own contradiction is not credible.
Afrimintel Ltd is independently owned, takes no operator, financier or sovereign funding, and publishes no ratings. No party named in this layer has reviewed, endorsed, commissioned or been informed of any record in it.
Where we have had contact with a company whose disclosures this layer examines, that contact is disclosed on the record for the asset. For the v0.1 seed: we have had research and outreach contact with the leadership of Lifezone Metals and previously delivered a sample asset dossier on Kabanga to the company. There is no commercial relationship, no payment in either direction, no subscription and no agreement of any kind.
It is disclosed because a reader examining a scrutiny product is entitled to weigh who its subjects are, not because it changed anything: every value in the seed is read from public documents, and the seed contains no finding against the claimant in the five fields searched — a bounded statement, not a clean bill on the asset. Queue selection does not consider counterparty status in either direction.
Any party named in a record carries a standing right of reply, published alongside the record and unedited. A demonstrated identity error, or a reconciling disclosure we failed to locate, is a correction under the error bounty. Corrections follow the published Quality Standard SLA: material within 7 business days, non-material within 30.
Afrimintel Ltd (2026). Afrimintel Divergence Record, version DR-0.1, published 31 August 2026. https://afrimintel.com/methodology/divergence-record/
Any citation of a record should state its classification and the dates of both sides.
Afrimintel · v2.137.88 · Divergence Record DR-0.1 — specification and evidence bar.