This is an experiment: how small and cheap can a model be and still spot reasoning problems in everyday text, such as politicians' speeches, company announcements, online forums, anything? If it can be made fast and cheap enough, it becomes realistic to run every piece of communication we care about through it, as a simple API. That's the classifier's job; the explanations you see here are for this page.
It runs on a tiny virtual machine in my closet, on CPU only, and the classifier still answers in about a tenth of a second. This is the baseline; a new model is in training.
How much to trust it
Treat it as a second opinion: it misses things, and some of what it flags is wrong. “How picky?” trades one for the other.
Your text
Texts aren't stored. Recent results are kept in memory for a while, so asking again is instant, and gone when the server restarts.
Developers: the API is at /api/v1 (classify, clarify, model); see /docs.