Supplementary Material: Partner Attribution Bias in LLM-Assisted Export Control Screen

This record contains the supplementary material for the paper "Partner Attribution Bias in LLM-Assisted Export Control Screening" (NLPAICS 2026), authored by Salem Alotaibi, Alexei Lisitsa, and Antony McCabe (University of Liverpool).

The supplementary PDF includes six appendices: the full experimental pipeline, the experimental logic diagram linking the four experiments, prompt structures and evaluation metrics, extended results including per-model embargo overlap and within-bloc country-level variation, illustrative rationale extracts showing how partner identity shapes model reasoning across geopolitical blocs and architectures, and geographic sensitivity visualisations for Experiments 1, 2, and 3.

All code, prompts, model outputs, analysis scripts, the UKRI EPSRC project descriptions, and the supplementary PDF are publicly available in this Zenodo record.

This work investigates the extent to which partner country identity influences export control classifications produced by large language models, using a counterfactual evaluation framework across five instruction-tuned models and 249 jurisdictions.