🤓 Latest Submissions
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.
13 Jul 2026
