"Consensus Complexity LawBench Public-Prediction Reanalysis"

"This dataset provides derived summary tables and reproducibility code for a secondary analysis of public LawBench zero-shot per-question predictions. The analysis measures model-output non-convergence among 14 large language models across 9 discrete or normalized-answer Chinese legal tasks, covering 1,800 legal items and 25,200 model-item predictions. The main metric, Delta_model, is defined as one minus the modal-answer share for each item. The package includes task-level divergence summaries, cluster-level summaries, full pairwise agreement tables, selected pairwise comparisons, task metadata, and Python code for reproducing the analysis from separately obtained LawBench public prediction files. The dataset does not contain raw LawBench predictions, new model outputs, human-subject data, or confidential legal materials. It is intended as a derived reproducibility package for research on legal AI evaluation, model-output divergence, consensus complexity, and semantic Byzantine fault. It should be interpreted as an operational proxy for evaluating model convergence behavior, not as a direct measurement of jurisdictional or institutional legal disagreement."