Safety-constrained and runtime-assured LLM multi-agent decision making for intelligent vehicle systems

Large language models offer flexible semantic reasoning for intelligent-vehicle interfaces, but probabilistic outputs, multi-agent disagreement, and provider failures complicate safety-sensitive use. We developed SCI-Proposed-v1.0, a lightweight multi-agent semantic decision architecture that combines transparent risk-informed routing with an independent Safety Validator and deterministic local Runtime Fallback. The system produces canonical semantic actions rather than trajectories or direct vehicle-control commands. In a frozen confirmatory benchmark of 150 unseen Base Scenarios, each represented by three reviewed language variants, SCI-Proposed completed 360/450 inputs (80.0%), compared with 169/450 (37.6%) for EI-MAS and 155/450 (34.4%) for a Single LLM. Language variants were treated as repeated observations within each Base Scenario, and all methods used the same frozen model configuration. Base-scenario paired differences were 0.4244 (95% confidence interval, 0.3489-0.5000; p < 0.001) and 0.4556 (95% confidence interval, 0.3800-0.5311; p < 0.001), respectively. Complete Safety Requirement Satisfaction was 316/360 (87.8%) versus 101/360 (28.1%) for EI-MAS. In frozen ablation, removing the Safety Validator reduced its prespecified endpoint by 0.2767 (p < 0.001), and removing Runtime Fallback reduced represented-failure recovery by 0.3667 (p = 0.00195). Exact three-of-three canonical action-set stability was 27/30 (90.0%). In prespecified isolated-sandbox hardware-in-the-loop validation, Scientific Task Success was 25/30, Safety Enforcement Success was 7/7 applicable cases, and Runtime Recovery was 4/6 applicable cases. Across these complementary, prespecified stages, separating generative reasoning from independent safety validation and class-bounded local runtime recovery materially improved semantic-decision dependability under the tested qwen-plus/DashScope configuration, frozen semantic policy, and represented fault classes. These findings do not establish broader vehicle-system safety, physical-control reliability, arbitrary-fault recovery, or deployment readiness.