Every World starts the same way. You point ARAIL at a subject and its research agents go to work — pulling papers, notes, manuals, whatever the subject is made of. A provenance gate throws out anything that cannot name where it came from. What survives is a corpus: every term sourced, every cross-reference resolving, every category declared. Clean, and yours.
The obvious thing to do with it is retrieval. Point an agent at the corpus, let it look things up, get grounded answers. That works, and it is what runs today. But retrieval means the knowledge lives outside the model, and every single question pays to haul some of it back in.
That is the end game. Everything the agents gathered and the gate cleaned becomes the training set for a small expert model of your own — one that does not look the subject up because it was taught it. Nucleus is the pipeline that does the teaching, and the model that comes out the other side is a 1–7B specialist you own outright.
THE PROBLEMSmall models are bad at reasoning
Here is the part that makes this hard, and the reason a naive version of this idea fails. When you shrink a model, the first thing you lose is not facts. It is reasoning.
Train a small student on question-and-answer pairs from a big teacher and you get a model that has memorised a lot of answers and cannot handle the question you did not think to ask. It pattern-matches. Move one variable and it falls over. Chain-of-thought — the step-by-step working that makes large models good at hard questions — was for a long time thought to be something only very large models could do at all.
So distillation has an obvious failure mode: you can compress what a model knows far more easily than how it thinks. A specialist that knows your domain but cannot reason inside it is a lookup table with extra steps.
THE FIXTrain on the reasoning, not the answer
Symbolic Chain-of-Thought Distillation is the technique that closes the gap (Li, Hessel, Yu, Ren, Chang and Choi, ACL 2023 — arXiv:2306.14050). Its finding is the one this whole pipeline leans on: students far below the scale where chain-of-thought was supposed to emerge — 125M to 1.3B in the paper — learn to reason step by step anyway, if you train them on sampled teacher rationales rather than on teacher answers.
So Nucleus never asks the teacher for an answer. It asks for the reasoning, then takes each rationale apart into explicit symbolic steps, and those chains — not the conclusions — become the training data.
Then the fine-tuning stage does something deliberately awkward: every training batch carries the right source documents and deliberate distractors. Plain supervised fine-tuning on clean pairs teaches a model to answer. Training against distractors teaches it to work out which ground actually supports the claim — so what it internalises is grounded reasoning rather than a reflex.
THE BAKESix stages, then a seal
The corpus goes in one end and a certified specialist comes out the other. Nothing in the middle is magic; each stage has a job you can name.
The adversarial stage is the one people find strange. Four agents spend their time trying to break the student you just trained — probing it, setting traps, scoring the damage, correcting it. It graduates when they stop being able to break it, not when a number of epochs runs out. Only then does an independent gate score it against benchmarks nothing in the loop ever touched.
The seal at the end records one thing, and it is worth being precise about what: where the model came from. The weights, the hash of the exact corpus bytes they were trained on, the configuration, the training record — bound together so the chain can be checked offline. That is provenance, not a grade. A seal tells you what a model is made of. It does not tell you the model is good; the gates before it are what argue that.
WHY IT MATTERSThe work was always the point
The reason this is the end game and not a side quest: you have already done the hard part. The months of research, the sources you checked, the corrections you made, the vocabulary you settled — that is the expensive input to training a domain expert, and most people never have it in a usable form. You do, because the gate made you.
So nothing has to be gathered twice. The World you can read is the World the model is taught from, byte for byte, and the hash that proves it is the same hash on both sides.
And then it runs where you are. QueueLLM streams the weights off SSD a layer at a time, so the machine you already have is enough — memory is the expensive, hard-to-source part, and disk is not. Frontier models in the cloud stay exactly what they should be: the teacher at build time, worth every cent for the reasoning they hand over. What changes is that the finished expert does not need one to answer.