Obsidia Router: Structure Before Inference

Most AI agents treat inference as the default: receive a request, call a model, spend tokens, then try to control the result. Obsidia reverses that order. Before inference, every request is compiled into a structured intent, evaluated by deterministic gates, and routed to the smallest admissible capability. For Track 1, Obsidia Router separates locally verifiable work from tasks that genuinely require Fireworks AI. Deterministic solvers handle supported structured cases, while ambiguous, unsupported, or risky requests are held, denied, clarified, or escalated. This reduces unnecessary model calls without replacing accuracy with aggressive guessing. The published evidence is explicitly scoped: 18/18 internal routing decisions were correct in dry evaluation, 740/740 invariants passed, and no false local closure was observed in frontier testing. In a live sample, four requests stayed local and one escalated, using approximately 389 Fireworks tokens. Hidden AMD evaluation results remain unknown until submission. For Track 3, the router is the visible edge of a larger idea: Obsidia as a cognitive operating system for governed AI. The LLM is an organ, not the sovereign decision-maker. Gates, receipts, bounded execution, memory constraints, and formal verification layers separate cognition, intention, decision, action, and proof. The principle is simple: structure before inference, and proof before irreversible action.