Backstage
The machinery, credited.
The library is the front hall; this is what builds it. Six pieces, each doing one job, none of them selling anything — the deck at /worlds is the proof they work.
ARAIL · the research lab
Autoresearch agents track the field daily; Buddy — a lab partner — watches your goals and flags what's worth attention. Local, zero telemetry.
Open the lab →DDaC · the press
Compiles a theme into a gate-checked World — a dictionary, an on-box Docent, and an open API at /dac. Every World in the deck, including Charlotte's, went through it.
The DDaC explainer →QueueLLM · the inference engine
Streams a model off SSD one layer at a time — load, compute, discard, prefetch — so the full weight set is never resident, with speculative decoding stacked on top. Apache-2.0; full credit to AirLLM for the layer-streaming idea.
Why & how →Nucleus · the bake pipeline
Seven stages — ingestion through 7-layer extraction, the adversarial swarm, three gates, a cryptographic seal — turn a teacher model and a gate-passed World into an owned, sealed specialist. Built end to end; today it runs deliberately, by hand.
Why & how →PaperAgents · the orchestrator
Small specialist agents, declared in TOML, reconciled to the state you declared when reality drifts — the shape the World-building swarm itself could take. pactl validate is its own contract today.
Codify a team →GeoAI · a proof in one real setting
A pre-production concept proving the stack end to end: vision-language reasoning over satellite imagery for humanitarian mapping, offline on volunteer hardware, zero-egress by design.
GeoAI overview →