To address the inherent tension between safety alignment and model utility in large language models (LLMs), researchers introduced CLEAR (Continuous Latent Adapter Routing). Unlike standard Supervised Fine-Tuning (SFT) or global Low-Rank Adaptation (LoRA) methods that apply safety constraints across all inputs, CLEAR utilizes a conditional adaptation framework. The method employs a lightweight hidden-state gate to continuously modulate the activation intensity of a safety-specific low-rank adapter. By keeping the backbone model frozen and dynamically routing through the adapter based on the input's latent representation, CLEAR minimizes performance degradation on benign prompts while effectively mitigating harmful outputs. Empirical evaluations on the Llama-3-8B-Instruct model demonstrate significant efficacy; the framework reduced the Attack Success Rate (ASR) on the HarmBench benchmark from 32.3% down to 0.5%. Crucially, this safety enhancement does not compromise reasoning capabilities, as CLEAR maintained most of the base model's utility and outperformed global tuning methods by up to 7.1 percentage points in GSM8K accuracy. These results indicate that continuous latent routing provides a superior mechanism for optimizing the safety-utility trade-off during LLM alignment compared to traditional, static tuning approaches.