Detect soft refusals in model output — the failures that look like success
Backed by a public CC0 dataset —
view or clone it.
Runs entirely in your browser.
FAQ
Why does this matter?
A refusal returns HTTP 200 with a polite paragraph. Automated pipelines score that as success and quietly ingest useless output. This catches it.
How is it detected?
Your text is matched against a public corpus of real refusal phrasings, categorised by refusal type. It runs entirely in your browser.
Does it work for every model?
The corpus covers common refusal patterns across major models. Phrasing varies, so treat the score as a signal — and open a PR if you hit one it misses.