Show HN: LitigationBench. A Litigation Task-Based AI Benchmark
Frontier and self-hosted models measured twice on the same litigation tasks: without Litco’s safeguards, and with them active inside the platform. Fabricated authorities, composites, and cost, published per model.
Summary not yet generated for this article.
Original Text
Adding to the sea of AI benchmarks, but with a focus on legal and litigation-based tasks and tests. E.g., hallucinating cases, misreading precedent, drafting, AI writing tics, etc.
Developed as part of a litigation platform I've started, but not sharing this benchmark to promote that. Just thought the findings were interesting--namely, Anthropic's models (except for Haiku) producing zero hallucinated cases.
Also, if anyone would like to see additional tests / benchmarks or models tested, let me know and I'll incorporate them into v3 if it makes sense.
Comments (0 of ) ⌛