Vons — Local Decisions

Vons — Local Decisions Compare candidate choices with a model on your device. You stay in control. Vons is a research-preview side panel for bounded decision experiments. Import your own compatible Vons ONNX bundle, enter context and a question, and compare 2–32 candidate options. Read the model proposal or abstention and its uncalibrated scores before deciding what happens next. MODEL REQUIRED No model weights are included or downloaded. You need a Vons full-graph Direct or Diffusion bundle containing bundle-manifest-v1.json, the ONNX graph, tokenizer and any declared external data. Shared/head-only graph partitions are not supported. General ONNX models are not interchangeable with a Vons bundle. The source and model-contract documentation are at https://huggingface.co/INLEVEL9/Vons. FEATURES - Local WASM/CPU inference; WebGPU is an explicit experimental option. - SHA-256 and file-size checks against the imported model manifest. - Clear abstention, input-budget errors and cancel controls. - Optional model retention on this device; Remove model clears the saved copy. - Response JSON, model provenance and single-run timings. PRIVACY AND CONTROL Only the sidePanel permission is requested. Vons does not read pages, browsing history or cookies. It does not execute actions, upload prompts, send analytics or require an account. Context and results are not retained after the panel closes. You choose whether to save the model locally. RESEARCH LIMITS This is not a general-purpose chatbot, autonomous agent or validated decision service. Scores are not calibrated probabilities of correctness. The research reports negative external-evaluation results and unresolved deployment checks. It is not intended for consequential medical, legal, financial or other high-stakes decisions. TERMS Qualifying noncommercial use is free under the Vons Community and Commercial License 1.0. Enterprise and other commercial use require a separate written agreement. Third-party dependencies and imported models retain their own terms. The research paper and Tech Report are available separately under CC BY 4.0. Contact: oswarld@inlevel9.com