Most assistants are optimized to always answer. Ask something outside what they know and they’ll produce a fluent, plausible, wrong paragraph — often with a citation that doesn’t exist. The confidence is the bug.
Abstain is a feature, not a failure
werbos is grounded in your library. When the answer is in there, it answers and cites the source. When it isn’t, it says so:
Not in your library — I won’t guess. I don’t know. Add the doc and ask again.
That sentence is worth more than a hundred confident guesses, because it’s the one thing you can actually build on. A system that sometimes bluffs can’t be trusted anywhere; a system that reliably abstains can be trusted everywhere it doesn’t.
Why this is hard
Getting a model to shut up is harder than getting it to talk. The default gradient pushes toward an answer. werbos treats the boundary of the library as a hard edge: retrieval either clears the bar or it doesn’t, and “doesn’t” routes to abstain rather than to improvisation.
The moat
Anyone can wrap a model. The hard, valuable part is the discipline: grounded answers with citations, and a clean “I don’t know” the moment you leave the grounded set. That discipline is the moat — and it’s the reason a small owned model can be more useful than a big rented one for the jobs that matter.