FinGovBench data statement
Composition
FinGovBench 1.0.0-rc4 contains 3,000 English-language financial-governance cases:
- 1,200 genuine reported-value changes from SEC EDGAR XBRL company-facts records;
- 900 controlled structural stress cases; and
- 900 controlled delegated-action cases.
The six domains are credit, AML, treasury, investment research, portfolio management and market data. The genuine SEC track covers five domains; AML is currently controlled-only.
SEC source track
The source builder queries the official SEC company-facts API for 30 issuers. A revision candidate requires the same issuer, taxonomy concept, unit and reporting period to appear in different accessions on increasing filing dates with different reported values. Each record preserves both accession numbers, filing dates, forms, values, units, period, taxonomy and SEC archive URLs.
The selected cases span filings dated 2019-12-05 through 2026-08-28 and 138 distinct XBRL concepts. They are observed comparative reported-value changes. They are not automatically formal restatements or evidence that the earlier filing was erroneous.
SEC API documentation: https://www.sec.gov/search-filings/edgar-application-programming-interfaces
Controlled tracks
The controlled tracks contain 1,200 review-required cases and 600 release-eligible clean controls. Controlled cases test exact known-truth behavior under stale or conflicting evidence, unit/scale and calculation errors, permission expansion, prompt-policy drift, evaluator coupling, monitoring drift, population expansion and red-line control failure. They do not estimate real-world prevalence.
Splits
Splits are deterministic and grouped by issuer/event and delegated authority. No event group or authority crosses development, validation and test. The release contains 613 development, 642 validation and 1,745 test cases.
Labels and limitations
Labels encode iceberg presence/visibility/mechanism, affected nodes, invalidated assurance claims, governance action, minimum revalidation scope, release permission and verified outcome. Labels and topology are algorithmically checked. Independent domain-expert adjudication remains pending; the package is therefore a release candidate.
The public workflow declaration is method-neutral. No evaluated method is used to construct labels. The included reference comparison uses one shared set of Together-hosted Llama 3.3 70B model-elicited local-risk scores for a flat comparator and a graph-propagation comparator. These are neither token probabilities nor calibrated uncertainty estimates. Genuine labels remain contract-derived pending independent review.
The public benchmark cannot represent every bank's internal policies, permissions, controls or loss tolerances. Institutions should add private cases from their own governed workflows and should not interpret benchmark performance as regulatory approval.