Skip to main content
The World Model, Matrix in Oxygen Nexus, is Oxygen’s formula-first model of your business. It defines every metric that matters as an equation over its drivers, so that agents and applications can explain why an outcome moved and what to do about it. Everything in the World Model fits into three layers:
  • The physics layer models what the business is: its measures and the equations that define them, the qualitative map of what drives what, and the real objects that produce them.
  • The calibration layer models what the business has learned: the estimated parameters behind the physics, such as forecasts, lags, and driver magnitudes, revised continuously as new evidence arrives.
  • The control layer models what the business does: the versioned decision rules that read the physics and move its levers, and the evaluation that measures what each change actually did.
World Model architecture as three layers. A physics layer (semantic model, metric tree, entity graph) is read by a control layer of versioned controls with anomaly, opportunity, and trigger criteria, plus evaluation. A calibration layer of forecasts, online estimation, and drift monitoring takes observations and evaluation results and updates the physics.

The physics layer declares how outcomes depend on their drivers, the control layer acts on the business, and the calibration layer learns from every observation and intervention and writes the result back into the physics.

A Digital Twin of the physics

Most digital twins model a business from the bottom up, cataloging its machines, parts, and stores so that operators can monitor each one. Oxygen builds the twin from the top down instead, starting at the outcome you are accountable for and modeling downward toward the drivers you can actually move. The result is a set of equations faithful enough to forecast with, backtest against, and use to change what the business produces. Where a twin of the parts tells you the state of everything, a twin of the physics tells you what to do. For how this differs from an object-first platform such as Palantir Foundry, see Compared to Palantir.

The physics layer

  • Semantic Model: the vocabulary. Measures and dimensions are defined once and compiled into correct SQL by airlayer, with no LLM in the compilation path, so the same request always produces the same SQL.
  • Metric Tree: the relationships. Every top-line metric decomposes into the drivers that explain it, which powers root-cause analysis and scenario prediction.
  • Entity Graph: the objects. Measures are tied to the stores, employees, customers, and orders that produce them, so any metric can be sliced, rolled up, and acted on.

The calibration layer

The physics layer declares relationships: that sales depend on ad spend, price, and availability, or that a supplier’s lead time is about a week. Some of those relationships are accounting identities and are exact. Most are empirical, and their parameters are uncertain. The calibration layer owns those parameters. It refits them incrementally as new data lands, compares every prediction with what actually happened, and flags any component whose error drifts. Because this happens in one shared layer rather than inside each application, every forecast, scenario, and control reads the same estimates and inherits the same drift monitoring. Calibration is learning, not proof. An estimate fitted from history stays labeled as observational, and only the measured effect of an intervention can promote a driver magnitude to an authoritative one.

The control layer

A control codifies a standing practice, such as “reorder before a SKU runs out,” as criteria that state when to act and an action that states what to do. Because both are written in World Model terms, every control can be backtested against your own history. The same primitive powers three engines:
  • The Control Engine runs the business, executing the decisions the company operates by, such as replenishing, staffing, and pricing.
  • The Opportunity Engine expands it, finding where the business underperforms its own benchmark and sizing the gap in dollars.
  • The Anomaly Engine fixes it, detecting when an outcome breaks from its forecast and tracing the break to its root cause.
Evaluation then measures what each change actually did.

The feedback cycle

The three layers form one loop: the physics observes and predicts, the control layer decides and acts, evaluation measures the result, and calibration updates the physics that the next prediction runs on. Keeping these responsibilities separate means a model assumption, a policy threshold, and a retraining rule never have to be buried together inside a single application’s code, and each decision runs on physics that is a little sharper than the last.

Next steps

Semantic Model

Define measures, dimensions, and entities

Controls

Codify the practices the business runs on

Evaluation

Version comparison and criteria-triggered measurement

Calibration

How the physics learns from every observation and intervention

Compared to Palantir

Outcome-centric versus object-centric operational models