Serve it live. Bake it permanent.
Part 1 served a fact live from the Worlds API. Below: the same corpus-hash lockfile that would pin a bake, applied to a second World — and the real memory math behind running whatever comes out the other side without a full GPU.
The lockfile a bake would pin.
Every World compiles to a corpus_sha256— pinned once, then echoed verbatim through hand-off, training, and seal. Part 1 hashed the Ecology World; below, the same mechanism runs on the Nucleus World's own dictionary — the World that documents this very pipeline.
the nucleus World, hashed
waiting for the corpus
same law, second World
The same assertClosedSourcedGraph from Act 03 — not a copy — run against this World instead.
waiting for the corpus…
the seven-stage pipeline · click a stage
Nucleus Seal
built the corpus-hash seal — deterministic, refusal-default, tested
aim the Ed25519-signed Nucleus Seal with live revocation
corpus
verified above
teacher
aim
config
aim
training
aim
chain_hash = sha256(corpus | teacher | config | training), signed once with Ed25519. The corpus half is what you just verified; the signing itself is the aim.
what the seal is designed to catch
Six things a signed chain hash defends against, each tied to which artifact it binds.
The four-artifact mint→verify path is implemented and tamper-tested on every signed field in the Nucleus engine. Hashing the audit log and AutoResearch report, embedding the seal into the model artifact, and the dynamic-monitoring loop are roadmap. Ed25519 proves what was signed — it is not, by itself, non-repudiation; production key custody is separate work.
However big it is, it runs bounded.
Nucleus bakes a settled World into a model's weights. QueueLLM is what runs that model afterward — streaming transformer layers off disk instead of holding the whole thing in memory. Pick a model below; the footprint is computed live, the same math the architecture page uses.
disk footprint · all weights
15.2 GB
Measured: 43 tok/s bf16, 3.13× mlx_lm (M5 Pro)
resident working set
8.2 GB
fits on…
the other half of "the World feeds the model"
A World doesn't only train a model — it can also constrain one at decode time. QueueLLM's grammar enforcer compiles a World's sourced vocabulary into a terms.gbnf grammar and masks disallowed tokens at the logit level, every step — not a request the model can ignore, a structural limit on what it can emit.
~31 tok/s
Qwen3-30B-A3B MoE
measured, 24GB Mac — real-time, no cloud GPU
400B+
Max model proven
streamed whole off a single 36GB Mac — minutes per token at this scale
~9×
Selective streaming
faster decode from loading only active experts, measured
You already ran a live model on this site — the ecology docent in Part 1 answered from a lookup slice, on-box, no cloud call. QueueLLM's architecture is what lets a model far larger than that run under a similarly bounded memory budget — today that's minutes per token at frontier scale, a proof of scale, not a chat.
While a topic is fresh you serve it as lookup — Part 1. Once it settles you bake it — the settled notes drop out of the active retrieval set, the model is the expert now, while the sealed compiled source stays for provenance and re-bake. The knowledge base compacts; the weights grow. Same World, kept current.
DEFINE (ARAIL) isn't covered by either demo yet — DEVELOP and SERVE ran live in Part 1; BAKE and COMPACT ran live just now.