Guard-Aware Trajectory Candidate Selection for Vision-Language Autonomous Driving: Safety-Constrained Chain-of-Causation

Driving VLMs use Chain-of-Causation (CoC) to link scenes to behavior, but CoC may omit when a trajectory is executable. We propose Safety-Constrained CoC, appending a natural-language guard with activation, invariant, bound, prohibition, release, fallback, and termination clauses. Given a scene, CoC, guard, and trajectory candidates, the model selects one candidate; an external deterministic verifier compares its measurements with the guard. Our paired-contract benchmark changes only the guard while fixing the scene, objective, and candidates. In three-seed temporal tests with LoRA-tuned Qwen3-VL-2B-Instruct, the violation rate decreased from 0.319 without a guard to 0.000, while compliant completion and paired accuracy reached 1.000. For unseen time values, candidate accuracy, violation rate, and paired accuracy were 0.635, 0.351, and 0.306. The results support guard-conditioned synthetic candidate selection, not continuous trajectory generation or vehicle-level safety.