ConsisGuard: aligning safety deliberation and policy enforcement in LLM guardrails

The ConsisGuard framework aligns safety deliberation and policy enforcement in reasoning-based LLM guardrails, bridging the "deliberation-execution gap" in which harmful intent identified in the model's reasoning is ultimately still judged safe. The framework uses policy-to-decision trajectory distillation and functional coupling alignment so that safety deliberation is grounded in safety policy and the final decision is derived from the deliberation. On prompt and response harmfulness detection benchmarks, ConsisGuard improves detection performance and reduces policy enforcement failures.