Large language models can understand passengers' driving commands, yet when acting as driving agents they reason only in generated text for a few seconds and must infer each action's consequences implicitly from scene coordinates. We introduce Jev‑AD, which couples Jev—a typed decision model that answers predefined questions with probabilities—not generated text—with a language world model that predicts how nearby vehicles move over the next two seconds for every candidate action. The code turns these predictions into foreseen consequences, which Jev compares directly instead of inferring them from raw coordinates.\n\nIn simulated highway merging (500 car episodes), this foresight reduces the same decision model's collision rate from 26.0% to 2.4% while taking 0.25 seconds per decision; Jev‑AD fulfills the passenger's intent in 93% of free‑form command trials.\n\nIn 500 replayed real‑world NGSIM episodes, foresight lowers the collision rate from 48.2% to 2.8%, and the world model learned from this traffic foresees 93% of dangerous actions. Project page: https://ys-qu.github.io/jev-ad-website.\n\nReview: Jev‑AD converts the vague reasoning of language models into verifiable safety constraints through explicit behavior prediction, achieving dramatic collision‑rate reductions and high command‑success rates, and thus offers a practical safety pathway for language‑guided autonomous driving.