General Unified World Model

A typed causal ontology of civilization, built on canvas-engineering structured latent spaces.

The General Unified World Model (GUWM) encodes 857 fields across 19 semantic layers into a single structured latent space. It learns from heterogeneous, partial data -- no dataset needs to cover every field. The model discovers cross-domain dynamics (how inflation drives bond yields, how policy shapes commodity markets) through shared latent structure and masked training.

All 857 fields packed onto a 128x128 canvas. Color = domain layer. Each pixel is a latent position with its own semantic identity.

Install

pip install general-unified-world-model

With data adapters and CogVideoX backbone:

pip install "general-unified-world-model[data,cogvideox]"

How it works

graph LR
    subgraph Data Sources
        A[FRED<br/>GDP, CPI, Rates]
        B[Yahoo Finance<br/>Equities, FX, Crypto]
        C[News<br/>Embeddings]
        D[Your Data<br/>Any CSV/API]
    end

    subgraph Canvas
        E["InputSpec/OutputSpec<br/>map columns to fields"]
        F["Canvas (T, H, W)<br/>857 fields on grid"]
    end

    subgraph Training
        G["CogVideoX Backbone<br/>(frozen, pretrained)"]
        H["Loop Embeddings<br/>(trainable, ~0.1%)"]
        I["Masked Loss<br/>only where data exists"]
    end

    A --> E
    B --> E
    C --> E
    D --> E
    E --> F
    F --> G
    G --> H
    H --> I

30-second quickstart

from general_unified_world_model import GeneralUnifiedWorldModel

# Create a world model for the domains you care about
model = GeneralUnifiedWorldModel(
    include=["financial", "country_us.macro", "regime", "forecasts"],
    d_model=64,
)

# Observe known values
model.observe("financial.yield_curves.ten_year", 4.25)
model.observe("country_us.macro.inflation.headline_cpi", 3.1)

# Predict everything else
predictions = model.predict()
print(predictions["forecasts.macro.recession_prob_3m"])

Or use the functional API for more control:

from general_unified_world_model import World, project
from general_unified_world_model.schema.business import Business

# project() accepts a schema root directly
bound = project(
    World(),
    include=["financial", "country_us.macro", "regime"],
    entities={"firm_AAPL": Business(), "firm_NVDA": Business()},
    d_model=64,
)

Visualizations

Geopolitical state — rotating globe

World model geopolitical state projected to RGB on a rotating globe. Each nation's color encodes a compressed vector of political stability, conflict risk, and economic alignment. Generated from mock data — trained model predictions coming soon.

Financial markets dashboard

Financial layer time series: yields, credit spreads, FX, equities, and volatility surface. The world model tracks all of these as simultaneous fields on a shared canvas.

Regime state dashboard

Regime latent: growth, inflation, financial cycle, credit cycle, liquidity, fragility, and systemic risk. The regime determines which causal channels are active.

RL environments from world models

Extract Gymnasium environments from trained world models. The same dynamics model yields different agent perspectives — employee navigation, CEO strategy, robot control — each with its own reward landscape.

env = model.to_openenv(
    obs_fields=["firm.financials.revenue", "regime.growth_regime"],
    act_fields=["firm.strategy.capital_allocation"],
    reward_fn=lambda obs, act, info: obs["firm.financials.revenue"].mean(),
)
obs, info = env.reset()
obs, reward, terminated, truncated, info = env.step(env.action_space.sample())

One world model → many RL environments. Each agent observes and controls different fields, producing different reward landscapes from the same dynamics.

See Environment Extraction for full docs, multi-agent support, and examples.

The 19 layers

Layer	Fields	Frequency	What it models
Physical	17	Annual+	Climate, disasters, geographic infrastructure
Resources	45	Hourly--Monthly	Energy, metals, food, water, compute
Financial	68	Sub-minute--Daily	Yields, credit, FX, equities, crypto, liquidity
Macro	67/country	Weekly--Quarterly	GDP, inflation, labor, fiscal, trade, housing
Political	42/country	Monthly--Multi-year	Executive, legislative, geopolitical, institutions
Narratives	35	Sub-minute--Monthly	Media, sentiment, elite consensus, positioning
Technology	13	Quarterly+	AI, biotech, quantum, robotics, productivity
Biology	16	Weekly--Annual	Ecosystems, disease, agricultural biology
Infrastructure	27	Hourly--Annual	Power grids, transport, telecoms, urban systems
Cyber	11	Daily--Quarterly	Threat landscape, digital ecosystem
Space	9	Weekly--Annual	Orbital environment, space economy
Health	10	Weekly--Annual	Healthcare capacity, public health
Education	11	Monthly--Annual	Education systems, workforce development
Demographics	10/country	Multi-year	Population, dependency, urbanization
Legal	11	Quarterly--Annual	Regulatory environment, rule of law
Sector	19/sector	Weekly--Quarterly	Per-GICS demand, supply, profitability
Supply Chain	9/node	Daily--Monthly	Bottleneck concentration, fragility
Business	57/firm	Daily--Quarterly	Financials, operations, strategy, risk
Individual	27/person	Daily--Quarterly	Cognition, incentives, network, state
Events	10	Sub-minute	News, filings, policy, conflict, disaster
Trust	17	Weekly--Quarterly	Data source reliability, epistemic state
Regime	17	Weekly--Decadal	Compressed world state, systemic risk
Interventions	13	Weekly--Quarterly	Policy actions, counterfactual effects
Forecasts	32	Output	Recession prob, credit stress, conflict

See the full Schema Reference for mermaid diagrams, field listings, and design rationale for every layer.

Status

Coming soon

The world model is under active development. The schema, projection system, training pipeline, and CogVideoX backbone are implemented and tested. Large-scale training on real data is in progress on H100 GPUs. Trained checkpoints and inference API are coming soon.

What works today: Schema compilation, projection, data adapters, heterogeneous training with masked loss, DAG curriculum, CogVideoX grafting, LLM-driven curriculum design, per-connection attention dispatch (17 attention types), dynamic layout/topology changes.

Coming soon: Pretrained checkpoint release, real-time data ingestion, hosted API.