← All benchmarks

SAB Research

Finance & Accounting

The research part of Mento.vc's Startup Agent Benchmark: 960 tasks in six venture-analyst workflows (product and customer, market and why-now, competition, team, traction and risks, sourcing screens), with 80 synthetic and 80 real-world tasks in each. An agent answers each task in a fixed JSON schema using only two task-scoped MCP tools, search and fetch, over a frozen corpus that is mostly distractors, and may answer unknown; facts removed from the corpus earn credit only for unknown. A deterministic grader gives up to 80 points for answer content and 20 for cited evidence, minus a penalty for nonexistent citations, and the score is the mean over tasks under the answer keys corrected after the authors' audit (v0.7.1). Answering unknown everywhere scores 7.5.

Top models (higher is better)