Obsidia Router: Structure Before Inference

Obsidia is a governed AI runtime built around one principle: Structure Before Inference. Most AI systems begin by deciding which model to call. Obsidia first decides whether a model is needed at all. Each request is translated into a UnifiedInputIR, converted into an Active Plan, matched against a Capability Resolver, and evaluated by deterministic governance gates before inference. The runtime follows this escalation path: Natural input → UnifiedInputIR → Active Plan → Capability Resolver → Gates → deterministic local solvers → readonly memory → optional readonly Brody endpoint → local Qwen inference → KX108_ONLY decision → canonical receipt → readonly replay. This architecture separates interpretation, planning, governance, inference, decision authority, and proof. Unsafe, destructive, incomplete, or locally solvable requests can be denied, held, clarified, or resolved before reaching a language model. When inference is required, a Qwen2.5-3B-Instruct Q4_K_M model is invoked as an organ of the system, not as its final authority. Every result remains governed by KX108_ONLY. The runtime records the selected capability, escalation level, gate verdict, invoked organ, model usage, mutations, external calls, latency, final status, and decision authority inside a canonical execution envelope. A SHA-256 receipt is generated, and readonly replay verifies the deterministic path without recalling the model or modifying memory. The Track 3 runtime passed 359/359 tests, including 5/5 Qwen tests. A 12-task demonstration batch produced 12/12 valid receipt hashes and 12/12 successful replays, with zero Fireworks calls, zero remote tokens, zero memory writes, zero world actions, and zero kernel mutations. Private Brody, Obsidure, Lean, Sigma, OIE, and domain components are not presented as executed. They remain private, route-only, unavailable, or bridge-required. Obsidia’s core idea is simple: the language model is an organ, not the brain, and never the final authority.