world-model-blueprint

Frames the autonomy objective, domain evidence, rights posture, model surface, data gaps, and adaptation path before downstream world-model work begins.

Plans Cosmos-class post-training, SFT, action modelling, checkpoint export, evaluation handoff, and serving readiness.

Plans production deployment, simulator integration, real-system shadowing, observability, incident response, rollback, and ownership.

Defines model evaluation, simulator tests, safety gates, regression suites, promotion evidence, and failure feedback loops.

Maintains governance, provenance, rights, license, privacy, export-control, safety, and audit evidence across the world-model lifecycle.

Selects and governs execution across local GPUs, Brev, OSMO, B200-class Slurm, Megatron, Kubernetes, Ray, and cloud launchers.

Routes neural reconstruction, Content Agents enrichment, harvested 3D assets, CAD-to-SimReady conversion, and asset-creation work to the correct NVIDIA/skills capability before SimReady handoff.

Plans serving through NIM, FlashDreams, Dynamo, Cosmos Framework inference, vLLM, Transformers, and edge or simulator runtimes.

Plans conversion and validation of CAD, OpenUSD, SimReady assets, Omniverse scenes, Isaac Sim scenarios, and synthetic-data pipelines.

Plans curation, enrichment, annotation, captioning, augmentation, split policy, and dataset release for world-model training and evaluation.