Financial research
FilingBench-50
All 50 original Finance Agent financial-research tasks, evaluated with DeepSeek using the original judge prompt.
About FilingBench-50
What it measures
Strict task success over all 50 source tasks; grader verdicts, execution errors and latency are recorded separately.
Scope & limitations
This result covers the complete 50-task upstream Finance Agent roster, not the four-task Blobfish companion world. One execution error remains ungraded and in the full-roster denominator. A single model run does not establish a cross-model ranking.
Evaluation environment
DeepSeek V4 Pro terminal scaffold in isolated Harbor environments on GKE. DeepSeek V4 Pro grades the original Finance Agent rubric.