From lurking and commenting... to a vacation that finally gave me time to chip away at the todo list my normal 60+ hour work weeks keep piling up.
I finally did the thing. I published a paper... my first public AI baby! yes, I'm a beaming new papa. lol
It is called Sovereign Knowledge Federation, and under the big name it is actually a pretty simple idea. Picture two people who each keep a really solid personal knowledge system, and each has an AI that can read theirs. One of them already cracked the exact thing the other is stuck on. Right now there is no clean way to hand that knowledge across. You can dump it on the open web, paste it into a chat, or give it to some platform that then owns it forever. None of those let the second person's AI actually use the first person's knowledge while still knowing whose it is and how far to trust it.
That missing piece is the whole paper: how independent people share slices of what they know, with nobody in the middle deciding what counts as true. Every claim carries a little signed tag for who said it, how sure they are, and where it came from, and your side gets to decide what to do with it.
Here is the part that genuinely surprised me, and it is my favorite bit to tell people. The obvious move is to label a sketchy source "low trust" and let the AI quietly discount it. I tested it on local models. It does not work. The model reads the label, sometimes even repeats it back like "noted, low trust," and then believes the claim anyway. So you cannot politely ask a model to be skeptical about something already sitting in front of it. Trust has to act like a bouncer at the door deciding what gets in at all, not a suggestion you whisper after it is already inside.
It is a working paper, which is the fancy way of saying it is real but I want it kicked around. So if you read it and something feels off, that is exactly the reaction I am hoping for. Whether this is your first month poking at local AI or you have been building this stuff for years, I want your take. Tell me where it breaks, and give me your ideas for the next phase.
Full paper here: