← All work
Python
modeldrift
A zero-dependency monitor for models you have pinned. It runs a fixed probe suite, reduces every answer to a behavioural fingerprint, compares that against a committed baseline, and fails CI when the model no longer behaves the way it did when you shipped.
modeldrift
Highlights
- Five aspects compared independently: format, structure, policy, length, lexicon
- Measures how much the model varies on its own, so noise is not reported as drift
- Suite is hashed into the baseline, so an edited prompt refuses to compare
- Offline stub backend, OpenAI-compatible and Anthropic endpoints, 94 tests
Want the code?
View on GitHub