Qorx Zero

Most AI agents send every request directly to a large model. Qorx Zero changes that order. It parses each evaluator request, identifies the task family, asks Qorx for a bounded local evidence pack, dereferences the evidence, and attempts an exact deterministic or strict local answer first. A model is used only when those routes cannot answer. Every response is checked against the evaluator schema before it is returned. ZeroRoute is written as a compiled .qorx program with a small Rust host, using the existing Qorx compiler, local index, context resolver, and proof-oriented evidence path. For Track 1, the container reads `/input/tasks.json`, writes `/output/results.json`, and runs on `linux/amd64`. Ordinary execution makes zero Fireworks calls. The optional Fireworks route is isolated behind explicit runtime controls and the evaluator-provided model list and proxy. The package runs without Jupyter or a local model, while its AMD AI Notebook supports the ROCm 7.2, PyTorch 2.9, and vLLM 0.16.0 environment for optional local inference. Telemetry keeps transport handles, resolved evidence, local-model input, provider tokens, and output tokens separate. The clean release passed 14 Rust tests, an eight-task contract test, secret scanning, archive extraction, Arch Linux WSL validation, and the container build. ZeroRoute is designed to preserve answer validation while avoiding model calls that are not necessary.