A world model for physical systems. Not a language model.
Graph Enabled Neural Intelligence Engine (GENIE) is a class of world foundation models for the autonomous operation of failure-intolerant physical systems; the first instance is industrial energy. Its job is state prediction — predicting the physical condition and dynamics of equipment (efficiency, degradation, fault trajectories, system response) from the signals it emits, not classification or raw-number forecasting.
It's physics-constrained and purpose-built. The architecture is closer to V-JEPA and DINO than to any language model — it predicts in latent space, forcing the model to learn physical structure rather than memorize patterns.
GENIE predicts in latent space rather than reconstructing raw signals — lineage of V-JEPA and DINO, not LLMs or diffusion. This forces the model to learn physical structure, not surface patterns. That's why its representations transfer to machines it's never seen.
Architected to integrate power signatures, IoT and edge-compute sensor streams, and operator knowledge. Each modality sharpens the state estimate. Full cross-modal fusion is the active research frontier.
Physics rules screen GENIE's answers. Each rule enters enforcement only after it passes validation on real instrumentation, against an acceptance bar fixed before the test, and it is enforced only on the inputs it was validated on. That's the prerequisite for acting in systems that can't fail.
What we've measured.
Every result below is on leakage-controlled evaluations against the strongest baselines available.
The V2.1 white paper reports what pretraining adds to equipment identification: the pretrained encoder reaches about twice the macro-F1 of an untrained encoder of the same design, in 5 of 5 training seeds.
Identifying equipment from raw power signatures, general-purpose AI scores near zero. GENIE — purpose-built — is near-perfect.
Physics-grounded anomaly detection transfers to unfamiliar equipment without retraining. AUROC 0.71 on held-out machines versus 0.57 for representation-similarity baselines. The model learned what degrading physical systems look like — independent of the specific machine.
Validated across multiple independent industrial datasets.
From interval-meter data — reliable where the monitored equipment is the majority of the load, the typical industrial profile.
Models initialized from GENIE's pretrained weights consistently outperform models trained from scratch on the same tasks.
One validated instance. A clear path.
An architecture that scales to every failure-intolerant physical system.
How a system responds to an intervention — the basis for true autonomous orchestration, gated on interventional data gathered through deployment.
The same world model architecture, re-instantiated beyond industrial energy — wherever failure-intolerant physical systems need a model that understands their physics.