← All benchmarks

Westworld Finance Diligence Bench

Professional Work · 2026-08-11

Halluminate's Westworld Finance Diligence Bench runs AI agents through 88 expert-authored tasks that together cover one company acquisition's due-diligence process (modeling, underwriting, quality of earnings, data preparation, pitch materials, closing and diligence review) on anonymized private-deal data, inside a Linux desktop with office applications, a data room, email and chat. Each task's grade is the weighted average of deterministic checks and agentic rubric verifiers, some with partial credit, so it lies between 0 and 1. Each model runs both in its developer's coding agent and in Halluminate's computer-use plus tool-use harness, which has no code execution, with three attempts per task. It is unrelated to Paper Instruments' DiligenceBench and to Diligence Stack. The score is the mean task grade over all 264 attempts, times 100.

Top models (higher is better)