RAG-LegalAI Data Compliance Benchmark and Study Materials

This dataset accompanies the manuscript “Retrieval-Augmented Legal AI for Data Compliance Decisions: Verifiable Evidence, Appropriate Reliance, and Decision Performance.” It contains (1) 245 deduplicated Main Benchmark tasks; (2) 295 source-governance evidence packets nested in 59 base tasks; (3) task-level scored outputs for the reported system conditions, together with aggregate, clustered-bootstrap, component, and weight-sensitivity results; (4) the frozen Study 2 stimuli, retrieval packets, questionnaire items, and provenance records; and (5) disclosure-controlled aggregate results from a three-condition randomized end-user experiment with 400 completed participants. Study 2 participant-level records, response-level records, event logs, reaction-time sequences, free-text feedback, session identifiers, participant codes, token hashes, timestamps, and consent hashes are not included. Public participant-derived tables apply a minimum cell size of 20; the smallest released task-condition cell contains 66 participants. Raw model prose is also excluded, while normalized predictions and all metric components used in the reported system results are retained. The controlled Study 2 database used to generate the aggregates has SHA-256 906c2280d3519b618f4db01c6669245974347878f6f497dc0c119eb4dd0d5da1 and is not part of this public release. Original research data and documentation are licensed under CC BY 4.0. Official legal texts and third-party services remain subject to their respective terms. The included legal-source files are researcher-prepared summaries and do not constitute legal advice.