LLMs are not databases. Meet System Zero.
Scott Wyatt · 2026-11-08
A language model predicts. It does not know. Squillo built the layer underneath that does, and gave it a number.
LLMs are not databases. So why is everyone treating them like one?
A language model does one thing, and does it astonishingly well. Given everything so far, it predicts what comes next. That is the whole machine. It has no record of what it told you yesterday, no file of what your team decided in March, no receipt for where an answer came from. When you ask it to remember, it does the only thing it can do. It predicts what a memory would look like.
Most of the industry has made the same bet: keep asking the predictor to be the memory, then bolt on enough retrieval, enough context and enough guardrails that the prediction usually lands. Usually.
We made a different bet. LLMs predict. Squillo preserves. The knowing is a separate system, it sits underneath the model, and we call it System Zero.
The wrong organ
Picture a team that has used an assistant for a year. Someone asks it: what did we decide about the launch plan in March?
The answer comes back in two seconds. It is confident, well written and plausible. It is also wrong, and nobody can tell, because there is nothing to check it against. The decision lived in a thread, a document and two people's heads. The assistant had seen fragments of all three and predicted the rest.
That is not a model failure. The model did exactly what it was built to do. The team asked the wrong organ. They asked the part that talks. They needed the part that knows.
Why zero
Your hand is off the stove before you have decided anything. Nobody voted. The body already knew.
That layer, the knowing that moves before thinking starts, is the one we built, and it is why the number is zero. It comes before System 1, the fast decision, and long before the slow argument. At Squillo the N Knowledge Model is System Zero and the N Decision Model is System 1. Both are built on N Lang, our language. The language model keeps the one job it is actually good at. It becomes the query language. You speak in plain words. It asks System Zero. System Zero answers, with a receipt.
What System Zero is
Three facts, in the terms of the knowledge model's own specification.
Knowledge is stored as functions, and the functions are experts. In a neural network, what the system knows is smeared across millions of floating-point weights. You cannot open it and read it. In System Zero a piece of knowledge is a function on the ledger: a typed N Lang program that can be called, composed with other functions, versioned, and read as plain JSON. Each one is an expert in the thing it knows. A question routes to the experts that own the facts, and the route is readable, not a learned black box. The explanation of an answer is the expert that produced it.
The ledger is the memory. System Zero is designed around an immutable ledger. Every function, every result and every change is an entry with a causal clock, so the order of events is provable without trusting anyone's wall clock. Fold the ledger to any point in history and you get the state as it was then. Even reading is recorded, so when the system inspects itself, the inspection is on the record too.
Nothing is thrown away, and everything travels. A transformer compresses everything it has seen into a hidden state of fixed size. System Zero's ledger keeps all of it. Reconstruction is selective, not lossy. That is what "every answer traces to its source" means mechanically: provenance complete by construction, the chain written while it runs, never reconstructed afterwards. And because an expert carries its provenance with it, it travels. It ships as a Snapp, a signed, versioned package of knowledge, and installs into another running model without stopping it. In The Matrix, Neo does not take a Kung Fu course. He gets an install. An expert installed for one skill sharpens that skill and leaves everything else exactly as it was. That is a theorem, not a hope.
The size of it
Experts are Snapps, so the whole thing is a registry, not a monolith. Semver for learned functions. Provenance on every piece of knowledge. A license on every expert. Nodes learn on their own and merge what they learned without a coordinator, so the system scales by adding nodes, not by buying a bigger machine. Knowledge composes like software because knowledge is software. Build it that way and there is no ceiling. It is a brain the size of the internet, growing by the expert.
Distilling it makes it smarter
Every memory system has to shrink. Keep everything at full size forever and it becomes unusable. The usual answer is to compress, and compression in a neural model means forgetting: the hidden state only holds so much, so something has to go.
System Zero's objective is built so that this trade does not exist. An expert's fitness is measured in bits per event, and that measure has a floor: the truth itself. The only way to reach the floor is to predict truthfully, so minimizing the loss and minimizing the error are the same act. Prediction and compression are one objective, not two. Distill a complicated expert into a simpler one with the same score and nothing it needs has been lost. It got smaller and truer in the same move.
Now turn that around, because distilling has a second edge. A frontier language model can be copied. Query it enough and you can train a cheaper model on its answers, because everything it is lives in weights, and weights leak through outputs. System Zero has no weights to pull. What it is lives in experts behind grants. A query gets exactly what its grant allows. A record you may not see leaves no trace in your answer, not even in the count. Whatever does come out carries its credits and its license. And every probe is itself an event on the ledger, and the ledger is what the experts evolve on. Try to distill System Zero and you are feeding it.
The decision
Every team using an assistant today is making one of two bets. Either the predictor is the memory, and you live with answers you cannot check. Or the knowing is its own system, with the receipt on every answer, experts you install like software, and a model that gets sharper the harder someone works to take it.
We made the second bet and gave it a number. If you want to see the language it is built on, N Lang is here. This is the first post on System Zero. The next one opens a knowledge function and reads it, line by line.
