Afrimintel · Platform · Decision Views · Sample dossier · Methodology
Ask about a covered African mineral asset, country, transaction, or safeguards profile. Answers are generated from Afrimintel's published data with the platform's epistemic discipline — every figure labelled Sourced / Derived / Absent, and nothing invented. If the answer isn't in coverage, it will say so rather than guess. It will not tell you what to do. Ask Afrimintel assembles the verified picture and names the gaps; it renders no verdict, rating or recommendation, and nothing it generates carries a Sourced tag — a model cannot certify its own source.
Ask Afrimintel is grounded in Afrimintel's published records and applies the platform's Quality Standard (Sourced / Derived / Absent) and conflict-of-interest scope. It does not offer investment, legal, or financial advice, and covers the Democratic Republic of the Congo, Rwanda, Burundi and Uganda on the same terms as every other jurisdiction, and states the jurisdiction interest disclosure those feeds carry when it answers from them. It renders no verdict, rating or recommendation, and characterises no party's conduct, in any jurisdiction — those rules are platform-wide, not regional. Generated answers can contain errors; verify any figure against the cited dossier before relying on it. Beta. Grounding coverage: 37 api/v1 feeds loaded on every question and two more loaded on demand (financeability, country production) — the full list is in ask-core.js. Every question carries today's date and the dataset as-of date, and every answer ends with that dataset stamp. When a question names one asset and a figure, the feed cell (value, state, source) is shown above the model's prose and quoted from the feed, not generated. Answer verification: every figure in a generated answer is traced back to the records above before you see it; any figure that does not trace is named and flagged, and the answer is marked not to be relied on. The check is deterministic code, not a second model. It is field-scoped: a figure must match the field it is claimed as, so a real value quoted under the wrong heading is caught, not waved through. A platform-wide count is checked against the census record rather than against the handful of records retrieved for the question — this is the class that produced a wrong answer on 10 September 2026, when the platform reported a total it had counted off its own sample. An operator or owner named in an answer must appear in the records. Its strength is measured rather than asserted: fabricated figures are run against real contexts at every deploy, and roughly 1.5 per cent of them still verify by coinciding with a real value, so a clean result is evidence, not proof. A fabricated claim carrying no digits is still not checked. The feed cells and the citation block are extracted by code and are not subject to any of this. Follow-up questions: a question that names no asset but refers back ("the project", "those terms") is read as a follow-up about the last asset you named, and the page tells you which asset it assumed so you can correct it. The carry lasts at most five questions, never overrides an asset you name, never applies to a platform-wide or methodology question, and is cleared by the "New subject" control. Nothing is retained between visits. Retrieval contract: tests/ask-retrieval-contract.js runs at every deploy and fails the build if any covered asset's own records fail to reach the model (check-85); tests/ask-answer-verification.js plants known-bad answers and fails the build if the verifier misses one (check-87). Measured accuracy: a 52-item eval battery is published at tests/ask-eval-battery-v1.json with an automated grader (tests/ask-battery-runner.js); no scored run has yet been published, so no pass rate is claimed here. That is a stated gap, not a good result.