Triduum Router

Triduum Router is a local hyperintelligence gateway and token-efficient reasoning proxy designed to extract deep analytical capabilities from open-weights LLMs like Gemma 4 on local hardware. Sits directly in front of a local inference engine (like vLLM), it serves as an OpenAI-compatible API gateway that introduces a dynamic routing scaffold, balancing execution speed and cognitive depth. When a query is received, the router initiates a lean token-streaming probe with logprobs enabled to determine confidence. By continuously calculating the margin between the top-1 and top-2 token predictions, the router measures model certainty in real-time. If the model is confident (exceeding the certainty threshold), the stream completes immediately, bypassing expensive reasoning pipelines to save substantial compute and time. For complex, ambiguous, or low-confidence prompts, the gateway halts the probe and triggers a Parallel CoT (Chain-of-Thought) Surge, spawning three concurrent reasoning pathways at temperatures 0.1 , 0.5 , and 0.8 . Each branch is guided by specific system prompts to focus on strict constraint validation, sub-problem breakdown, or creative devil's advocacy. These pathways are then evaluated in parallel by three distinct consensus agents: • The Logical Constraint Auditor to filter path failures against rules. • The Sequential State Tracker to verify chronological consistency and logic flow. • The Syntax & Format Compiler to enforce structural correctness. Finally, the Core Synthesis Engine combines these evaluation vectors to produce a unified, high-fidelity response. Designed for local deployments, Triduum Router is optimized for AMD ROCm GPUs, offering developers a plug-and-play agentic reasoning endpoint with zero client-side integration changes.