Why closed-wall
The AI doesn't get to guess.
Every general-purpose AI is built to always have an answer. We built ours to refuse when it shouldn't.
What hallucination actually is
A language model trained to predict the next likely word will always produce one. When it does not know, it produces a plausible word anyway. That is the failure mode.
How closed-wall prevents it
We constrain the model to only read from a vetted library. If retrieval comes up empty, the model is instructed to say so, not to fill the gap with training data or open-web fragments.
What we will never do
- Scrape the open web to find an answer.
- Let the model freelance when the library is silent.
- Present a guess as a fact.
- Cite a source we cannot actually point at.
Common questions
What does closed-wall mean?
The AI is only allowed to read from our curated library. It is not allowed to fall back to the open web, training data, or anything else when answering you.
What happens if the library does not cover my question?
We tell you. We do not guess. A 'we do not know yet' is a better answer than a confident wrong one.
Doesn't that limit what you can answer?
Yes, on purpose. Coverage is growing every week. We would rather refuse 10% of questions than confidently mislead on 100%.
Isn't this just RAG?
Retrieval-augmented generation is part of it, but closed-wall is a stricter rule: the model has no fallback. Most RAG systems still let the model freelance when retrieval comes up short. We don't.