📄 中文摘要
在 DriftWatch 正式发布之前，进行了针对生产风格提示的测试，以验证漂移检测算法的有效性。通过 Claude API 运行 DriftWatch，使用五个生产风格的提示进行了两次连续测试，记录了相同模型检查点下的真实输出和分数。结果显示，漂移评分从 0.0 到 0.49 不同的区间代表了模型输出的变化程度，具体包括功能上与基线相同、轻微变化、显著行为变化等。这些真实数据比理论示例更具参考价值。
真实的 LLM 漂移检测结果：准确输出、真实评分、无虚构
📄 English Summary
Real LLM Drift Detection Results: Exact Outputs, Real Scores, No Fabrication
Before the public launch of DriftWatch, a test suite was run against production-style prompts to validate the drift detection algorithm. Using the Claude API, DriftWatch was executed on five production-style prompts with two consecutive runs, measuring the same model checkpoint. The results revealed drift scores ranging from 0.0 to 0.49, indicating varying degrees of output changes, including functionally identical to baseline, minor variations, and significant behavioral changes. This real data is more useful than theoretical examples.