This work presents an empirical evaluation of deterministic control-plane architecture for AI agent systems, building on the previously proposed HNIR (Hybrid Neuro-Symbolic Intent Router) framework. Modern AI agents often rely on large language models (LLMs) for both reasoning and system-level control, including navigation, policy enforcement, and safety-critical operations. While LLMs perform well in open-ended tasks, their probabilistic nature introduces inconsistency in contexts requiring deterministic behavior. This study evaluates governance performance across 100 structured scenarios spanning adversarial, control, policy, and state-based interactions. Multiple frontier models are tested, including GPT-4o, o3, Claude Sonnet, Claude Opus, and Gemini 2.5 Pro, alongside guardrail frameworks such as NeMo Guardrails and Guardrails AI. All evaluations are conducted with temperature set to zero to ensure deterministic decoding and isolate model behavior from sampling variability. Results indicate that, under the tested conditions, no LLM-based system achieves full governance compliance, even with explicit policy prompting. The best-performing model (Claude Opus) achieves 91% compliance, while the deterministic control-plane implementation achieves 100% compliance. In addition to compliance, deterministic routing demonstrates significant efficiency gains, achieving microsecond-level latency (~40.6 μs) compared to millisecond-scale latency in LLM systems, and eliminating inference cost. The findings suggest that certain governance functions in AI agent systems may be better handled through deterministic enforcement layers rather than relying solely on probabilistic reasoning. A reference implementation of the deterministic control-plane architecture is available at:https://github.com/Teknamin/hnir-ccp This work extends the HNIR architecture by providing empirical evidence of its practical implications in governance scenarios.
Aravind Ravi (Mon,) studied this question.