Diagnosing Live Within-Policy Instruction Conflicts in LLM Agents with Witnessed Resolution Profiles

Researchers introduce WIRE, a diagnostic pipeline for detecting conflicting rules within LLM agent prompt policies. Testing six public policies, the system identified 170 rule-pair conflicts and found that 64.6% of witnessed conflict scenarios resulted in at least one source-rule violation, revealing significant gaps in how language models handle competing policy directives.

LLM agents increasingly operate under complex, multi-rule prompt policies designed to govern behavior across diverse scenarios. These policies are typically written in natural language by different authors at different times, creating the risk that individually sensible rules can conflict when applied simultaneously. WIRE addresses this gap by systematically extracting rules from policies, encoding them as testable clauses, identifying hard collisions through satisfiability checking, and measuring how models actually resolve conflicts in practice.

The research reveals a critical reliability problem: across six major public policies, only 35.4% of tested conflict scenarios achieved joint compliance with both competing rules. This 64.6% violation rate signals that deployed LLM agents may systematically fail to uphold their stated governance policies when rules interact. The methodology is rigorous—extracting 276 rules, encoding 560 atomic clauses, and testing 13,335 scenarios—suggesting the findings reflect genuine policy brittleness rather than noise.

For AI developers and enterprises deploying LLM agents, this work exposes a blind spot in policy validation. Organizations typically audit policies in isolation but rarely test how rules behave under realistic conflict conditions. The distinct resolution patterns observed across different models and tool actions indicate that conflict resolution is not deterministic, adding another layer of unpredictability. As LLM agents take on higher-stakes decision-making roles in customer service, finance, and content moderation, understanding and fixing these conflicts becomes essential. The WIRE framework provides a practical starting point for auditing existing policies and designing more robust policy systems.

• →WIRE identified 170 within-policy rule conflicts across six major LLM agent prompt policies, with 64.6% of conflict scenarios violating at least one source rule.
• →Only 35.4% of tested conflict situations achieved joint compliance with both competing policy directives.
• →The methodology combines rule extraction, satisfiability encoding, and witness realization to surface hidden policy conflicts before deployment.
• →Different models and tool-action types show distinct conflict resolution patterns, indicating unpredictable behavior under policy pressure.
• →Policy conflict diagnosis is now actionable for AI teams but does not measure real-world deployment frequency of these conflicts.