Validation of autonomous driving (AD) functions demands diverse, safety-critical scenarios that are both semantically faithful to natural language requirements and executable in standards-compliant simulators. This paper presents LLM4AV, a text-to-simulation pipeline that converts open-ended descriptions into a domain DSL and renders ASAM OpenSCENARIO 1.x on OpenDRIVE road geometry selected via rule-based map retrieval. A multi-layer validator enforces XML syntax, kinematic feasibility, and behavioral-logic consistency (e.g., maneuver timing, trigger coordination). Closed-loop executions are performed in esmini, and safety is quantified using a segment-scoped protocol that evaluates metrics only during maneuver-valid intervals to avoid dilution from staging or cruising. Across eight scenarios spanning highway maneuvers, urban right-of-way, vulnerable road users, and low-friction braking, LLM4AV consistently produces executable scripts with high semantic fidelity. Observed risk profiles align with domain expectations, including a worst-case TTC = 0.30 s for a bicycle-overtake near an occluder, PET = 0.15 s for an occluded mid-block pedestrian crossing, and peak longitudinal jerk 8.52 m/s3 under wet-road hard braking; benign cases (e.g., emergency-vehicle yielding) maintained large headways despite high Δv. The contributions are: (i) a standards-aligned NL → DSL → OpenSCENARIO synthesis chain, (ii) a practical OpenDRIVE retrieval method with feasibility repair, and (iii) a criticality-aware evaluation protocol that concentrates metrics on high-risk windows. The results indicate that LLM4AV improves reproducibility and interpretability of scenario testing while preserving portability across standards-compliant simulators.