Pin the Envelope, Not the Model Name
说明与用途
使用三种摘要的三重检测器可发现免费/试用端点中的隐藏模型更改,防止评估分数回归。
问题
Regression from routing changes in free/trial endpoints can silently alter evaluation scores.
结果
Tripwire detects silent model swaps by comparing request, envelope, and stamp digests, preventing unnoticed regression in evaluation scores.
工作流
- 1. Generate request fingerprint of the prompt and parameters.
- 2. Compute envelope digest of the response text, finish reason, and usage keys.
- 3. Record declared stamp (endpoint identity).
- 4. Store all three digests per run in a JSON line file.
- 5. On subsequent runs, compare request, envelope, and stamp digests to baseline.
- 6. If request digest unchanged but envelope digest differs and stamp unchanged, flag a potential silent model swap.
- 7. Use paired sign test to assess statistical significance of score changes.
证据
Every call should produce three independent digests, because each one answers a different question about who changed what.
Write one JSON line per case per run, and never overwrite the file.
- 记录日期: 2026-10-16T06:40:59.347Z