Analyst
Ask a question about the FinObservatory estate: banking, currency and sovereign-debt crises, US bank health, systemic risk, financial conditions, sovereign debt, AML/CFT reference data, and EU and Bangladesh banking. The analyst answers by querying the datasets, so every figure it gives traces to a cited primary source. It follows the StatGPT pattern: the model turns your question into a structured query and never invents a number. How it works.
Nothing here is graded by assertion. A published harness re-runs 123 questions against ground truth it recomputes from the data at run time. The latest full live pass, 2026-07-19, scored 122 of 123, and the question it missed is named, with the scoring rules, on the methodology page linked above.
Answers are computed from the FinObservatory data estate and cited to their primary source. The analyst never generates numbers from memory: every figure comes from a query against the datasets. A quote from the document library carries a locator into its source document, the character offset and span, plus the paragraph where the corpus stores paragraph breaks, so the passage can be checked against the original. It refuses questions the data cannot answer, and it gives no investment advice and no predictions.