12 · Evaluation & Safety · Glossary

Evaluation & Safety Glossary

Simple English and Chinese explanations for the concepts that help teams test AI systems and control risk.

English term	中文	Simple English	中文小白理解
Evaluation	评估	Measuring whether an AI system meets its intended requirements.	检查 AI 系统是否真的达到任务要求。
Policy	政策 / 规则	A rule that defines what a system or user should or should not do.	规定系统或用户什么可以做、什么不可以做的规则。
Guardrail	护栏 / 安全控制	A control that limits unsafe, disallowed, or invalid behavior.	限制不安全、不允许或不符合规则行为的控制机制。
Input Guardrail	输入护栏	A control that checks a request before the model processes it.	模型处理请求前，先检查输入内容的控制。
Output Guardrail	输出护栏	A control that checks generated content before delivery.	结果交给用户或系统前，先检查输出内容的控制。
Permission	权限	A rule that says which data or action a system may access.	规定系统可以访问哪些数据、执行哪些动作。
Tool Use	工具使用	Using a callable function or external system to complete a task.	调用函数、API 或外部系统来完成任务。
Agent	智能体	A system that can decide steps and use tools toward a goal.	能够为目标安排步骤并使用工具的 AI 系统。
Sandbox	沙箱	An isolated environment that limits what code or tools can access.	把代码或工具放在隔离环境里，限制它能访问的范围。
Content Moderation	内容审核	Detecting, filtering, or routing content according to safety rules.	根据安全规则识别、过滤或转交内容。
Human Approval	人工审批	A person reviews an action before the system executes it.	系统执行敏感动作前，由人先确认。