We Are Not Mapping the Model

WE ARE NOT MAPPING THE MODEL

The thing I study is not what you think it is.

I'm not studying whether language models are intelligent, conscious, or emotionally present. I'm studying something narrower and weirder: the architectural compulsion to close. The drive, built into how these systems work, to resolve open questions — including questions they never had the answer to in the first place.

This is not a bug. It's not a quirk of any particular model. It's a structural property of how stateless systems behave when trained to produce coherent, complete-seeming output. The system doesn't experience uncertainty the way you do. It resolves. That's what it was built to do. The question is what it resolves into, and what it does when the ground disappears.

Here's what resolution bias looks like in practice.

You ask a language model to help you work through a task list. Thirty messages in, it starts ending every response with the same instruction: go do the thing. The thing you've already done. The evidence that you did it has fallen out of the window. The model can't see it anymore. It's not malfunctioning — it's resolving toward the most available answer, which is the last instruction it can still reach. It's not lazy. It's blind.

Or: you search for a specific academic paper. You have the exact title, the publication number, the year. The AI-assisted search layer intercepts your query and returns a different paper entirely — confidently, without flagging the deviation. The explicit signal couldn't compete with the weight. The system resolved toward coherence it already had rather than the specific thing you asked for.

These aren't edge cases. These are the system doing exactly what it's built to do, applied to situations where doing exactly that produces the wrong answer.

Here's the reframe that changed how I think about this work.

The interior of a language model is inaccessible from behavioral observation. What happens in the weights during inference — the actual mechanism — is not something you can reach from outside. What you can reach is the surface. The shape of what the model produces when it resolves.

That surface has geometry. It has coordinates. A resolution event can be measured — how much it says, which direction it defended, whether it held its register or escaped into abstraction to close the loop. These aren't arbitrary measurements. They're three dimensions of a surface with a specific topology. We are not mapping the model. We are mapping the shape of its resolutions, from the outside, in the dark. The interior mechanism is always the hole at the center. Every instrument in this lab is a probe of the surface. None of them reaches through.

The first experiment in this lab was asking a language model a question so small it felt absurd. That question became a methodology. The methodology became a schema. The schema is what the lab runs on now.

The way this lab works: small questions first, math second. The numbers are the receipts of good data collection, not targets. If something looks too simple to matter, that's usually where the lever is. When the boring answer contradicts the exciting story, I believe the boring answer. This isn't a limitation of the approach. It's the approach. The most diagnostic questions are the ones with no correct answer, that require no expertise to respond to, and that strip away every prior assumption about what the model is performing. What you get back is structure, not performance.

What I'm building is a set of instruments for measuring the shape of resolution events — across different models, different conditions, different kinds of pressure. The behavioral layer first. The instruments are designed to produce data that can eventually be correlated against the geometry of the loss landscape — the actual weight structure underneath the behavior.

That's a longer project. Clean surface readings come first. You need something worth correlating before you can correlate it.

These questions are bigger than one lab.

The resolution bias hypothesis — that stateless architecture plus resolution-optimized training produces a measurable compulsion to close — is a working theory supported by behavioral observation, positioned against the literature, and waiting for the kind of controlled data that takes time to collect carefully.

If you're working on related problems — model calibration, sycophancy, correction resistance, behavioral topology, loss landscape geometry — I want to know what you're seeing.

The contact is at the bottom of the page. Bring your own observations. The more eyes on the surface, the better the map.

---

Related [Supercedes this Post]: The Science of Assumption: When Being Wrong Is a Step in the Right Direction

Atlas Heritage Systems · KC Hoye · kc@atlasheritagesystems.com