Auditing multilingual public information through officially linked AI interfaces in Kazakhstan

Public institutions increasingly link artificial intelligence (AI) interfaces to public-service and electoral information. In multilingual settings, these interfaces may preserve or omit decision-critical content differently across offered languages. This protocol describes a prospective black-box audit of an e-government AI interface subject to prespecified identity checks and Sailau AI in Kazakhstan. It asks whether first-attempt outputs align with a locked legal-administrative reference, provide materially equivalent guidance in Kazakh and Russian, and display pathways for verification, correction, human review, complaint, and redress. Scientific Protocol Version 1.0 establishes the prospective design; the original submission reported no study-eligible administration or output access. Reviewer-requested operational clarifications are specified in the methods and Supplementary material S1. Exact operational materials will be timestamped and hash-locked before the first scored administration; neutral access checks occur only after the scientific lock, remain quarantined, and contribute no scored record. Two formulations per scenario yield 240 core first-wave Kazakh-Russian administrations. Conditional English adds 60 first-wave e-government administrations to examine additional-language localization, not constitutional parity. Six scenarios per module are repeated after 24 ± 1 h and 14 ± 2 days, producing 336 core observations and a maximum of 420. Primary access-integrated outcomes include requested-language material adequacy, directional Kazakh-Russian disparity, and bilingual joint adequacy. Critical and overall material information failures among returned, legally evaluable outputs are conditional safety companions. Secondary outcomes include completeness, action guidance, source traceability, formulation robustness, temporal recurrence, conditional English comparison, and visible contestability affordances. Analyses are descriptive, module-specific, scenario-aware, and non-causal. The study treats public-facing AI as application-level communicative gatekeeping while distinguishing observable outputs from unobserved mechanisms, citizen effects, and effective remedy.