Develop a lightweight automated red-teaming tool that injects epistemic pressure into model prompts to detect sycophancy. This could be integrated into LLM evaluation pipelines to harden system instructions.