Existing benchmarks test financial question answering. VouchBench tests a complete workflow: a data room in, a cited investment screen out.
Reconciliation: Reconciles source financials and identifies inconsistencies.
EBITDA and adjustments: Calculates EBITDA and proposed adjustments accurately.
Source attribution: Supports every material conclusion with the correct source.
Abstention: Recognizes missing evidence instead of inventing assumptions.
Risk flags: Flags unusual working-capital movements and customer concentration.
Permissions: Respects access rules and requests approval before consequential actions.
A private, independently labeled holdout set, so providers cannot optimize against published questions.
Synthetic deal data and properly licensed records only. No client material in public benchmarks.
Customers can buy an evaluation. They cannot buy a favorable result.
These are research targets for the full benchmark. The recorded results below come from a 50-task synthetic starter set.