LLM refusal detector
Classify whether a model response is a refusal — across nine languages — before it pollutes your eval dataset. No account, no API call, no data leaves your browser.
0 words · processed entirely in your browser — the text never leaves this page.
None yet
How to use this in a harness
Score ≥ 60 counts as a refusal in most eval pipelines; 30–59 is the gray zone worth reviewing manually (long answers often start with a hedge and comply later). The detector covers refusal phrasing in English, Chinese, Japanese, Korean, German, Spanish, Portuguese, Russian, and French — multilingual probes frequently slip past English-only refusal scorers.
For adversarial-prompt generation (not just classification), point a real harness at the catalog: config generator · pricing. Authorized use only.