Ask the agent a question and watch what it does with one it cannot answer. It runs on GoldenWall, a fictional manufacturer's corpus, so nothing here is a real product and nothing you type reaches a real customer's documents.
Most demos are built to show a system at its best. This one is more useful pointed at its edges, because the refusal is the part that decides whether a technical department can put its name on it.
It is answering from GoldenWall, a fictional manufacturer's document set. We use an invented corpus rather than a customer's so that a public demo never puts a real client's literature, or a real client's gaps, on the open internet.
Nothing you type is used to train a model. Questions asked here are logged against the fictional corpus so we can see what it could not answer, which is the same mechanism a client gets.
A fictional manufacturer called GoldenWall, built specifically so a demo can be public without putting a real client's literature on the open internet. A client deployment runs on that client's own corpus.
Because the answer is not in the corpus. That is the product working, not failing. Try asking it to interpret a building code and watch it decline and offer to route the question instead.
Same engine, same grounding, same refusal logic, same citation behaviour. What differs is the corpus and the channels it is connected to.