| # Method independence |
|
|
| FinGovBench evaluates governance of a financial AI workflow. It does not require any |
| particular model architecture or graph-based submission. |
|
|
| ## Independent benchmark object |
|
|
| Each public case declares the operational workflow boundary, versioned components, |
| dependencies, evidence available at the case's `as_of` time, assurance claims, authority |
| tier, permitted actions and observed change. This is a system specification analogous to |
| an architecture or lineage record. It contains no method-specific probabilities, path |
| contributions or certificates. |
|
|
| Gold outcomes are defined by the financial event and governance contract: |
|
|
| - SEC cases use an official accession-level reported-value pair and a declared workflow |
| contract; their labels remain provisional until independent domain adjudication. |
| - controlled cases use a prespecified event, authority boundary and executable contract. |
| - the evaluator, not a submitted method, determines whether the selected action and |
| revalidation scope permit release. |
|
|
| ## What a submission must do |
|
|
| A method receives the same workflow declaration and must return the affected |
| components, governance action, revalidation components and release decision. A submission |
| may use rules, a flat uncertainty model, an LLM, a knowledge graph or another method. |
| It may transform the neutral workflow declaration into its preferred internal format. |
|
|
| ## Optional graph-based methods |
|
|
| Graph-based uncertainty propagation is one optional evaluated approach. It must read |
| exactly the same public closed-loop inputs as lexical and flat-uncertainty comparators. No |
| method-specific telemetry or labels are provided. |
|
|
| ## Anti-circularity requirements |
|
|
| Before version 1.0: |
|
|
| 1. independent reviewers must adjudicate event validity, required action and revalidation |
| scope without seeing any method prediction; |
| 2. submitted predictions cannot alter gold labels or case inclusion; |
| 3. all methods receive the same public inputs and sealed test labels; |
| 4. genuine and controlled results remain separate; and |
| 5. any threshold or learned policy freezes on development data before test evaluation. |
|
|
