Research Data for “Beyond Document Retrieval: Bitemporal Regulatory State Reconstruction for LLM-Based Trade Compliance”

<p dir="ltr">This dataset contains the machine-readable research materials supporting the study “Beyond Document Retrieval: Bitemporal Regulatory State Reconstruction for LLM-Based Trade Compliance.”</p><p dir="ltr">The dataset includes frozen evaluation cases and Gold benchmark targets, E/H/F evidence packages, authenticated prospective model outputs, parsed predictions, analysis data, frozen evaluation protocols, parser and scoring materials, runtime and failure records, and a legal-source manifest.

The evaluation covers U.S. Section 301 tariff classification, EU Carbon Border Adjustment Mechanism (CBAM) regulatory-state classification, and CBAM liability calculation across two model families.</p><p dir="ltr">The three evidence conditions are: E, a flat retrieved-evidence package; H, an integrated structured regulatory-state package combining source lineage, information time, legal-effective time, scope, and domain state; and F, a frozen operative-rule reference package.</p><p dir="ltr">The package is intended to support transparent reproduction of the reported benchmark-internal results, including full-denominator performance, response-production and parseability decomposition, liability exactness measures, and post-hoc sensitivity analyses.