World model
A world model is a learned model of how an environment responds to actions: given a state and an action, it predicts the next state. For the web as an environment, DOM event streams can contribute observed state transitions by placing interactions and later page changes on the same timeline.
How it works
World models support planning by predicting likely next states for candidate actions. Training one for web interfaces requires observations of states, actions, and later states in a usable format. A DOM event stream places interactions and page mutations on the same timeline, which can make candidate transitions easier to extract than they are from video. Ordering and timestamps show sequence, not causality. Network responses, timers, concurrent scripts, and other actors can change the page after an action. A training pipeline therefore needs attribution rules, environment metadata, and filtering before treating recorded intervals as action-result pairs. Held-out recordings can then supply observed transitions for evaluation.
Where you encounter it
Model-based RL literature, agent planning research, and infrastructure arguments about what behavioral data is for.