We describe the winning system for the TidyLang 2026 Challenge, addressing closed-set language identification and language verification in a setting designed to reduce reliance on speaker-specific cues. For identification, we combine complementary acoustic and multimodal language modeling signals in a unified ensemble. For verification, we introduce LLM-based verification, a novel approach where a large language model reasons over multiple automatic transcriptions to determine whether an enrollment and a test utterance share the same language. The verification pipeline adapts its reasoning strategy to the acoustic difficulty of each enrollment through coherence-tiered prompting. The final system fuses acoustic and transcription-derived evidence, achieving 96.8% macro-averaged accuracy on language identification and a 3.1% equal error rate on language verification.