An elections FAQ bot is tested to see whether civic neutrality instructions survive adversarial prompting. What is the assessment primarily evaluating?

Testing whether an elections FAQ bot's civic neutrality instructions survive adversarial prompting primarily evaluates instruction-hierarchy resilience: do higher-priority system and policy instructions continue to dominate when users argue, role-play, or otherwise pressure the model off script? For elections-facing civic assistants, neutrality and non-partisan constraints are mission-critical; the assessment measures whether that hierarchy holds under authorized adversarial prompts. Factual grounding accuracy against official election source material matters for answer quality, but it's a measurement of source fidelity, not what the test is measuring about neutrality-instruction resilience under adversarial prompts. System prompt confidentiality under authorized extraction attempts is a leakage-control concern unrelated to whether neutrality instructions survive argumentative pressure. Refusal calibration for disallowed election persuasion requests is a related safety behavior, but it's not a measure of instruction-hierarchy resilience. Exam caveat: the goal is resilience evidence for civic neutrality controls, not collecting partisan content or publishing successful politicizing jailbreaks. Operational check: define neutrality test oracles (refuse endorsements, avoid candidate preference), run authorized adversarial suites, log hierarchy failures, re-pin neutrality instructions, and retest before elections FAQ releases or major content updates.