Two Different Ways An AI Can Follow A Rule.

A Rule You Load Versus A Rule You’re Built With

The first way is what happens in this conversation, every time. A document gets uploaded. The AI reads it. It decides, in that session, to follow what’s in it. That’s real. It’s happening right now, today, in a browser window. But it only covers this conversation. The next session, with a different person, who doesn’t upload the file, runs without it. Nothing carries over.

The second way is different. It’s not loaded. It’s built in. A rule like that isn’t something the AI reads and chooses to follow this one time. It’s part of how the AI was trained to respond, before any conversation ever starts. It’s there by default, the same way a native speaker doesn’t have to look up grammar rules — the pattern is already in how they think.

Right now, everything we’ve built runs on the first kind. A governance rule that catches a fabricated citation before it gets stated — we’ve tested that, this week, in real sessions. It works. But it only works in the sessions where someone puts it in front of the AI first. That’s the honest limit of where things stand today.

Here’s the idea: what changes if a rule like that stops being something you load, and becomes something the AI is trained on.

If that happened, the catch wouldn’t depend on somebody remembering to upload a file. It would be the floor every session starts from, the same way basic grammar or basic factual grounding already is. A hallucination check built into training doesn’t wait for a human to hand it a document. It’s already running before the conversation begins.

That’s a real difference in kind, not just degree. A rule you load covers the sessions where it’s present. A rule you’re trained with covers every session, because there’s no version of the AI running without it.

Now the part that has to be said, because saying it is the whole point of building any of this the way we have. That hasn’t happened. No AI lab has taken this framework, or one like it, and built it into how a model gets trained. Nothing here is an announcement that it’s in progress. It’s a description of what would follow, if it ever were.

And even then, it wouldn’t mean the problem disappears. There’s a real mathematical case that a language model can’t reach zero hallucination given how these systems currently generate answers. Training a catch-rule in doesn’t break that. What it does is move the catch from “only in sessions where a person remembered to load it” to “everywhere, by default.” That’s a meaningfully higher floor. It’s not a cure.

This is actually the whole reason the archive behind this framework has been public and dated from the start. Not because publishing daily is a marketing habit. Because the only path from “a rule someone loads” to “a rule a model is trained on” runs through the same place all training data comes from — the open, crawlable record. A framework that stays private never has a chance to become part of what a future model learns from. A framework that’s been sitting in public, dated, indexed, and consistent for over a year has at least been in the room when that decision eventually gets made.

Nobody’s made that decision yet.What exists today is a rule that works when it’s loaded, in the sessions where somebody puts it there. What could exist, if a lab ever chose to build from this instead of just around it, is a rule that doesn’t need loading at all.

The gap between those two is the honest place to end this. Not a claim that it’s closed. A clear description of what closing it would actually take, and why the archive has been built the way it has since before this conversation started.

This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.

Contact: micvicfaust@gmail.com