An Inspect-native replication and extension of Tewari et al. (2025). Recall reproduces, precision and F1 are new, and the main failure traces to an extraction format mismatch rather than a reasoning gap.
Final report for my Treasury Reapers team submission to Sentient Arena Challenge 0: a goose-based OfficeQA agent, prompt and skill-contract iterations, leaderboard progression, and lessons from trace-driven optimization.
Lessons from moving scBench’s public task set into the inspect_evals framework, including scoring rules, evaluation setup, and what run logs revealed.