MODEL.SURGERY JOINS XERJ:
HACKERS AND MATHEMATICIANS BUILDING NEUROSYMBOLIC AI.
A research team of AI hackers and mathematicians, Model.Surgery, has decided to join XERJ. Together we will build neurosymbolic systems: models that propose, programs you can read, and facts that check the answer. We have bigger plans than a single model. Both teams want to make something that has never existed before.
Why they came, and why we said yes
Model.Surgery is a research team of two kinds of people who like to work on the same object. The hackers operate on weights. They cut, split and graft a network in place, without a new training run. The mathematicians bring deep experience in the theory of topology, and look for the shape of what a model has learned.
They chose XERJ because the engine already runs a symbolic layer. You or your agent write facts. XERJ Query-1 turns a question into a short program. The engine runs that program over those facts. Every answer comes from a fact, with the words it used. With no fact, the engine says UNKNOWN. With two facts that disagree, it says CONFLICT. You can read the program, and you can run it again.
That same layer is what they saw in XERJ Reader-1, now in preview: a model that writes a formula over a document, and an engine that calculates the value from the words on the page.
I haven't seen anything this new in AI in a while that didn't come from China. Let's crash it together, here in the US.
That is a place where a team that hacks on weights, and a team that thinks in shapes, can do work that neither could do alone. From today we write as one. Our shared aim is AI that does the work without overspending.
Small, verifiable models for enterprise data
On 10 November 2026, Anthropic, PBC filed Amendment No. 2 to its Form S-1 with the U.S. Securities and Exchange Commission (File No. 333-299633). The registration statement has been filed and has not yet become effective. In the Risk Factors section it writes:
Moreover, demand for frontier models may decline if customers determine that less advanced, lower-cost models are sufficient for their particular use cases. As AI capabilities improve across the industry and a broader range of models becomes available, customers may conclude that incremental performance gains at the frontier do not justify the additional cost, and may migrate toward smaller, cheaper, or open-source alternatives that adequately address their needs.
That is exactly where we are focused: small, verifiable models for enterprise data. This is work without overkill: a model sized for the job, and an answer you can check. XERJ Reader-1, now in preview, is built for this work. It writes a formula that points at the exact words of a document, and a separate engine calculates the value, so every value can be checked.
The Reader-1 announcement shows it at work on documents like that. Asked how many days after a power purchase amendment was signed the new price begins, Reader-1 wrote a formula over two dates at opposite ends of the text, and the engine calculated 161. We charge $0.01 per 1M input tokens, with output free. In our measurements, a question about a short document, about a third of a page, came back in about 1.2 to 1.6 seconds. Every approved account starts with a $1.00 free balance.
What is real, and what the mathematics cannot yet explain
Transformers work. That is real. Scale them and they get better, in ways you can measure. Show them a few examples in a prompt and they pick up the pattern. At enough scale, new abilities appear that nobody wrote down as a rule.
The mathematics has not yet caught up with that practice. We do not have a complete account of why these models generalize to so much they were not trained on. We do not yet know, as geometry, what their internal representations are. We do not yet know why a short prompt can rewrite their behavior so deeply. These are open questions, widely acknowledged. The systems have moved ahead of the theory.
We decided not to wait for the theory to arrive on its own. We will combine practical hacking, and some ideas that still sound unlikely, with deep mathematics, especially the theory of topology, and try to move the horizon of frontier theory to where the systems already are.
Neurosymbolic execution
A model is good at proposing. It is a poor place to hide an answer. In our vision the model writes a short program, the engine runs that program over facts, and the facts, with the words they came from, are what you check.
That is neurosymbolic execution: a neural part that reads and proposes, a symbolic part that executes and checks. Query-1 already works this way over facts in XERJ. XERJ Reader-1, now in preview, extends the same idea to documents, with a formula that points at the words of a page. Surgery gives us a way to reshape the model itself, on your side of the wall, without a new training run.
We want to keep doing that kind of work: change the model, keep the operation visible, and let the mathematics catch up in public.
Move the horizon
We want to change how AI is built. A model that shows its work can be checked. A model you can operate on can be studied. We want that work to contribute to new science. If the practice of transformers has run ahead of the theory, then this is how we try to close the distance: practical hacking, bold ideas, and the mathematics of shape, pointed at the systems that already work.
The new models will come from that join. We will share them as we make them. This post is the start of that work, not a list of names or dates.
One prompt for your agent
Set up XERJ for me, following https://xerj.org/llms.txt. Install XERJ with the XERJ Query-1 model and start it. Index the folder I point you to (ask before reading anything outside it) and write three facts from that data. Ask XERJ Query-1 one question about them, and show me the answer, the program and the evidence.
The vision begins with tools already in your hands: a local engine, facts you wrote, and a program you can read. The new models will follow.