Financial AI evaluation

Evaluate a research system against a defined piece of work, not a general promise of accuracy. A filing question, a multi-document report, and a recurring monitoring task have different inputs and failure modes. This collection starts with those distinctions and connects benchmark interpretation to checks you can run on your own permitted sources.

Keep documented product features separate from results you have actually measured. Freeze the task, source set, scoring rules, and product version before comparing outputs. Preserve incorrect answers, missing evidence, and review time alongside successful examples so the result explains where the system needs human intervention.