World simulation
Agents and physics
The Sim skill turns Nexma into a 3D agent-based simulation platform. It models agents, environments, physics, and interactions, then runs them at scale to predict emergent behavior. Where the other skills design a static network, this one presses play on it — the world model becomes a dynamic system you can scrub through time.
What this skill models
The ontology captures a simulation world with the entity types, link types, and constraints a modeler expects.
- Entity types.
Agent,Environment,Object,Force,Collision, andPhysicsBody. Each carries typed properties — mass, velocity, perception range, behavior policy. - Link types.
AgentMovement,EnvironmentalInteraction, andPhysicsEvent. Events carry timestamp, participants, and outcome state. - Key constraints. Physics determinism per timestep; collision resolution under configurable solvers; perception-radius limits; and agent-count budgets per worker.
The ontology ships the reference data a modeler relies on: standard physics-engine defaults, gravity and friction tables, and behavior-tree primitives.
What Jax can do
Jax reasons about populations and scenarios because the Skill binds its capabilities to this ontology.
- Generate populations that match a target distribution across age, behavior class, and capability. "Simulate 5,000 pedestrians evacuating this stadium under alarm conditions. Use exit-prefer behavior with a 12-meter perception radius."
- Branch scenarios — run N variants with parameter sweeps and surface the policy that hits the goal.
- Detect bottlenecks in pedestrian, vehicle, or material-flow simulations.
- Explain causally — trace back through agent state history to explain why an outcome happened.
- Validate conservation checks, deterministic replay, and a bottleneck report.
Scenario sweeps and large runs dispatch through the Nexma MathEngine; trajectories and aggregate metrics write back to the world model so you can scrub the timeline the moment a run completes.
Tip: Confirm a deterministic replay before running a parameter sweep. If two runs of the same seed diverge, the sweep is comparing noise, so validating determinism first makes every branch comparison meaningful.
Constraints and standards
The skill enforces the rules a simulation must satisfy to be trustworthy.
| Concern | Rule enforced | Standard |
|---|---|---|
| Determinism | Same seed, same result | Per-timestep replay |
| Conservation | Mass and momentum preserved | Physical conservation laws |
| Pedestrian flow | Social-force dynamics | Helbing social force |
| Collision avoidance | Reciprocal velocity obstacles | ORCA |
| Scale budget | Agent count per worker | Runtime capacity |
Export targets include trajectory CSV, video capture, and aggregate metrics from Project to Export.
Out of scope today: continuum fluid simulation (agent-based only), photoreal rendering (visualization is functional, not cinematic), and multi-million-agent runs in a single browser session — a server backend is required at that scale.
Where to go next
- Defense and military — rehearse a C2 plan as a scenario.
- Logistics — material-flow simulation feeds routing decisions.
- Skills overview — the full catalog.
- Nexma MathEngine — the dispatch layer behind large simulation runs.