#reasoning

Why reasoning and tool calls leak into answers when self-serving open-weight LLMs on vLLM or SGLang, classified by layer, with engine settings, client contracts, and a remediation layer that prevent it