How it works · Chapter 6
Models propose, rules decide
Why the step that decides where a case goes is not an AI model at all.
Every AI step in Hamsa produces a structured answer: a claim, a match score, a label, a confidence, a list of sources. None of them decides what happens next. A short set of written rules reads those answers and decides, in the same order, every time.
Why a quick check needs an unfakeable signal
A model's confidence is not enough to put a case one click from publishing, because a model can be sure and wrong. A quick check needs something independent of the model: the frames matching footage from somewhere else, an AI generator's watermark, a published fact-check that agrees, or two independent trusted sources, each on top of a confident draft. Everything else goes to a checker, with the reason it fell short written on the case.
When the evidence conflicts (a source that supports the claim against one that refutes it, or a Nepali translation that does not match the English), the case always goes to a person, with the conflict shown.