# Polaris (rust-norion)

A Rust research prototype for a DNA-inspired, self-evolving control layer around LLM inference, handling routing, memory, reflection, and device adaptation around an existing inference backend rather than being a production inference engine itself.

rust-norion, called 北极星 (Polaris), is a research prototype and control-layer engine for LLM inference, built for non-commercial research use. It focuses on the layer around inference rather than the inference kernel itself: routing, memory, reflection, device adaptation, and backend boundaries, explicitly stating that it is not a production-ready inference kernel, not a wrapper around closed models such as Gemma, Llama, or Qwen, and not a way to bypass GPL or other model and data licenses or maintainer review.

It runs as a Cargo workspace of crates: norion-core for the control-layer abstractions, routing, KV handling, and hardware profiles; norion-memory for memory, gist, experience, indexing, and governance; norion-agent for agent workflow, task assignment, and reflection; and norion-cli, currently a no-backend protocol shell for testing CLI and TUI routing without starting a real model. It supports built-in local heuristic backends, an external command-runtime boundary, and a manifest-gated production-runtime boundary for loading real model assets, plus a device-profile and execution-plan gate.

Self-evolution capabilities are described as preview, gate, and rollback oriented, favoring evidence, audit, and write gates over freely allowing persistent mutation.

Website: https://github.com/yanghao1143/rust-norion