Hallucination Detector
How the Hallucination Detector Works
You are the last line of fact-checking against a model that is confidently, eloquently, and frequently wrong. Statements stream across your screen as cards. The model swears every single one is true. The model is not to be trusted.
- Cards drift across the screen — each is a short "fact"
- Some are true. Some are plausible hallucinations the model invented
- Tap a card to flag it as a hallucination and deny it
- Let true facts scroll off the edge untouched
How You Lose Trust
Flag a real fact as fake, or let a hallucination escape to production, and you lose a life and your combo resets to x1. Three strikes and the model ships unsupervised. Catch lies in a row to ramp your combo multiplier and your score. Spawn rate and drift speed climb the longer you survive — the model gets more confident, not more correct.
Slop Fact: Large models don't "lie" — they sample the most probable next token, which is sometimes a citation to a paper that was never written by an author who does not exist. We call this "creativity" in the deck and "a bug" in the postmortem.