Verdic – 人工智能系统意图治理层
1 分•作者: kundan_s__r•6 个月前
我们在生产环境中部署 LLM 时，反复遇到相同的问题，因此构建了 Verdic：大多数 AI 故障并非源于内容安全，而是意图漂移。
随着模型变得更具自主性，输出结果常常悄无声息地从描述性行为转变为指令性行为——没有任何明确的信号表明系统正在采取行动。在这种情况下，关键词过滤器和基于规则的防护措施很快就会失效。
Verdic 是一个意图治理层，位于模型和应用程序之间。它不检查主题或关键词，而是评估：
* 输出是否将未来的选择压缩为特定的行动方案
* 响应是否施加规范性压力（指导行为 vs 解释说明）
目标不是内容审核，而是行为控制：检测 AI 系统何时在其部署的意图之外运行，尤其是在受监管或决策关键的工作流程中。
Verdic 目前以 API 的形式运行，具有可配置的允许/警告/阻止结果。我们正在对自主性工作流程和长期运行的链进行测试，这些场景中意图漂移最难检测。
这是一个早期版本。我主要希望收到在生产环境中部署 LLM 的用户的反馈，特别是在以下方面：
* 自主性系统
* AI 治理
* 风险与合规
* 我们可能遗漏的故障模式
欢迎提问或分享有关该方法的更多细节。
查看原文
We built Verdic after repeatedly running into the same issue while deploying LLMs in production:
most AI failures aren’t about content safety, they’re about intent drift.<p>As models become more agentic, outputs often shift quietly from descriptive to prescriptive behavior — without any explicit signal that the system is now effectively taking action. Keyword filters and rule-based guardrails break down quickly in these cases.<p>Verdic is an intent governance layer that sits between the model and the application. Instead of checking topics or keywords, it evaluates:<p>whether an output collapses future choices into a specific course of action<p>whether the response exerts normative pressure (directing behavior vs explaining)<p>The goal isn’t moderation, but behavioral control: detecting when an AI system is operating outside the intent it was deployed for, especially in regulated or decision-critical workflows.<p>Verdic currently runs as an API with configurable allow / warn / block outcomes. We’re testing it on agentic workflows and long-running chains where intent drift is hardest to detect.<p>This is an early release. I’m mainly looking for feedback from people deploying LLMs in production, especially around:<p>agentic systems<p>AI governance<p>risk & compliance<p>failure modes we might be missing<p>Happy to answer questions or share more details about the approach.