PSEBench - Clause Accuracy: leaderboard

Metric: Clause accuracy (%, M2): the targeted reportable clause is correct, on cases where both the gold and the predicted verdict are Reportable, on the 5,074-case MN29 benchmark (3,455 complete, 1,362 missing-information and 257 uncertain cases built from clause cards of the Minnesota 29 reportable adverse health events), agentic environment where the model may ASK a GPT-5.2 information provider within 10 turns and then answers with a verdict, clause, legal evidence and rationale; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.

Top models

#ModelScore
1GPT-5.598.5
2GPT-598
3Claude Sonnet 4.697.9
4Gemini 3.1 Pro (Preview)97.8
5Claude Opus 4.797.3
6Gemini 2.5 Flash96.7
7DeepSeek R195.3
8GPT-OSS-120B94.3
9Qwen 3 235B A22B 2507 Instruct93.3
10Mistral Small 3.289.5
11GPT-5 Nano87.2
12Llama 3.1 8B Instruct59.7

Interactive version: theaggregate.ai/benchmark?slug=psebench-clause-accuracy · How It Works · Data refreshed daily, snapshot 2026-09-29.