Data minimisation guards in digital identity wallets: benchmark, code and model outputs

Benchmark, necessity policy, experiment and analysis code and all stored model answers of a study of on-device language-model guards that flag unnecessary attributes in OpenID for Verifiable Presentations (DCQL) requests to a European Digital Identity Wallet. The archive contains 768 clean, 180 permissible-attribute and 630 attacked requests with English and Turkish purpose texts, the outputs of seven open-weight models as free judgment and hybrid guards, an adaptive attacker, follow-up runs (a fixed rule for unknown purposes under direct attack, a second attacker model, repeat runs and a larger classifier) and an independent coding sample. All purpose and attack texts are synthetic. README.md describes the layout and how to reproduce the tables.