Orchestration for Genie-class action-conditioned world models (Phase 176.5). The
trained dynamics plug in through the WorldDynamics interface; a deterministic
ReferenceWorldDynamics (a real navigable closed-form world) makes the whole
stack runnable and testable without a GPU.
WorldSession— the durable, resumable session: deterministic seed, prompt log, pose-keyed@nous/wm-memoryMemoryBank, input/frame history, mid-session re-prompting, a transport boundary (WebRTC video + input, fail-loud when absent), and a serialisable durable state tree.GenieWorldService— the/maya/genie/createendpoint: text prompt → interactive world session, persisted to aSessionStore(in-memory default; MinIO in 176.8.1.2), with resume-after-eviction. (176.5.1.5)AdversarialPromptController— the "adversarially prompt your experience" PvP mode: multiple players submit world-rewrite prompts, arbitrated by priority- per-player cooldown, broadcast through a
@maya/multiplayerbinding. (176.5.2.2)
- per-player cooldown, broadcast through a
EducationScene— the@metis/lessonswrapper: a teacher prompt plus learning objectives + assessment hooks, tracking real exploration state (regions visited, prompts issued, pose) and emitting progress reports. (176.5.2.3)- Trainable video stack (
nous_genie/) — exact-default 8×16×16 VQ-VAE tokenizer with a 16,384-entry EMA codebook, causal eight-code latent action model, and action/prompt/persistent-memory-conditioned autoregressive dynamics. The full 11B recipe is represented without allocating it in local gates.
Not included (external — left unchecked in TODOS)#
The large-dataset training, published-quality interactive rollouts, and ≥24 FPS
720p B200 result in 176.5.1.1–176.5.1.4 require large-scale accelerated
training on hundreds of thousands of video hours. Their locally actionable
architectures and conditioning paths are implemented above.
text
npx vitest run --root libs/nous/genie
cd libs/nous/genie && python3 -m pytest tests-python/ -q