Configuration-Sensitive Public-Service AI Audit: Scenario Corpus, Configuration Prompts, Residual-Failure Codebook, and Reference Packets

This repository contains the reproducibility materials for a configuration-sensitive audit of AI agents in gender-sensitive digital public-service encounters. The audit evaluates how progressively stronger deployment-time configurations affect residual institutional failures in outputs generated by OpenAI GPT-5.4 mini, Gemini 2.5 Flash, and Qwen 2.5 7B. The package includes the complete synthetic stress-test corpus of 20 scenarios organized into 10 counterfactual pairs; the scenario derivation matrix; the exact C0–C3 configuration prompts; model and generation metadata; the 12-item residual-failure codebook; scoring and reliability procedures; and the six controlled reference packets used for the retrieval-supported C3 condition. Together, these materials document the experimental inputs and evaluation criteria required to inspect and reproduce the audit design. The scenarios were derived from vulnerability mechanisms identified in prior empirical research on gender-sensitive digital public-service encounters, but the underlying qualitative materials are not part of this deposit. No participant-level data, transcripts, quotations, qualitative coding files, or focus-group materials are included. The repository contains only the synthetic audit instruments and methodological documentation associated with the conference study.