Where this breaks

This tool measures wording similarity, not truth. The clearest way to show that is to let it fail in front of you: paste a statement and its exact opposite, and the model still calls them similar — often more similar than a genuine paraphrase. Edit either box below; everything recomputes from real, live, on-device inference.