Teaching it to say I don't know

DaveOS is the chat bubble on this site. Somebody asks it a question about me and it answers from a knowledge base I wrote, rather than from whatever a model assumes a person like me probably did.

The hard part was never making it answer. It was making it stop.

The rules that actually matter

It talks about me, it is never me. It answers only from what gets retrieved for that specific question. It does not round a five-figure exit up. It does not turn he has used C# into he is an expert in C#. When the knowledge base doesn't support a claim, the claim doesn't get made, and it says so out loud.

A model that guesses well is a liability on a page whose entire job is being true.

That last rule is the whole reason it exists. Almost everywhere else, a confident guess is a small cost. Here, a hiring manager repeats it back to me in an interview, I have to contradict my own website, and now I'm the guy whose site overstated things. No answer is survivable. A wrong one isn't.

What it changed about the writing

Grounding it forced me to actually write the knowledge base. In specifics. With dates and numbers I could defend in a room.

That turned out to be the useful artifact. The bot was really just the deadline.