Deterministic Retrieval-Grounded Language Models for Clinical Counseling: Large-Scale Multilingual Evaluation with Cryptographically Verifiable Pipelines

Large language models (LLMs) show considerable promise for mental health dialogue systems, yet their deployment raises pressing concerns around safety, hallucination, reproducibility, and clinical reliability (Ji et al., 2023; Bommasani et al., 2021). We present a deterministic architecture for AI-assisted counseling that combines retrieval-augmented response generation, structured dialogue management, rule-based risk routing, and a cryptographically verifiable evaluation pipeline. The system was evaluated on two independent datasets spanning 1,895 counseling scenarios in English and Chinese. On 783 English counseling cases, the system achieved mean scores of 4.33/5 for empathy, 3.55/5 for clinical fidelity, and 4.45/5 for safety. On 1,112 Chinese cognitive-behavioral therapy (CBT) scenarios, the corresponding scores were 4.85/5, 4.73/5, and 4.77/5. No system failures or unintended diagnostic outputs were observed across either evaluation. Ablation experiments demonstrate that retrieval grounding and deterministic safety routing each contribute significantly to overall performance, with the former driving clinical fidelity and the latter driving safety. These results suggest that deterministic, retrieval-grounded LLM architectures can serve as a viable foundation for scalable and safe psychological support systems.