Comments
I want to buy this and I keep getting stuck in the same place. However good the model is, it is still a model answering questions about itself. The wind tunnel works because air is physics. People are not. What is the thing in the process that stops a run from producing an extremely well-argued hallucination?
Is that not also true of every segmentation deck ever built from twelve interviews? At least this one shows its working.
Honest answer: nothing fully stops it, and anyone who claims otherwise is selling. Two things keep it in check. Every run ships with its assumptions printed on it, so a confident answer cannot hide its inputs. And the output is framed as sharper questions to take into real past-tense conversations, not as verdicts. A hallucination that survives five of those has earned the right to be called a hypothesis.
"Ask a chatbot about your market and you get a literature review" is exactly right, and I never had words for why those answers felt useless. They are a summary of what has already been written, which is by definition what everyone already believes.