World model

The world model is the system's internal simulation of how its environment responds to actions. Before the arbitration layer commits to a candidate proposal, the world-model agent evaluates it: given the current state and this proposed action, what outcome is likely, how confident is the model in that prediction, and what is the risk if the prediction is wrong? These estimates feed directly into the arbitration scoring - predicted reward is 30% of the arbitration score, and the confidence estimate modulates how heavily that predicted reward is weighted. The world model's accuracy starts low and improves through the prediction-error feedback loop with the grounding engine: every time the world model makes a prediction and the subsequent telemetry reveals the actual outcome, the grounding engine computes prediction accuracy and routes it back to update the model's confidence calibration. Domains where the model is consistently wrong or consistently uncertain become targets for the curiosity engine. The dream engine uses the world model as its evaluation substrate: synthetic scenarios are generated and then evaluated against the world model's predictions, which is only useful if the world model is already reasonably calibrated - another reason the integrative developmental stage, which builds world-model calibration, must precede the open-ended stage.