Open Bias releases LLM policy proxy

Open Bias is a proxy that sits between LLM applications and model providers, reading RULES.md to evaluate requests and responses in real time.

• It applies pending interventions in pre-call, and runs sync/async engines in post-call.
• Violations are handled in three ways: block, intervene, shadow.
• It provides a default starter RULES.md and an automatic evaluator, and lets you adjust the engine, tracing, and enforcement method via openbias.yaml.
• Engines are divided into judge, nemo, fsm, and llm; judge is async so core-path latency is 0ms, nemo is 200-800ms, and fsm and llm are experimental.
• Non-critical violations are processed asynchronously to be reflected in the next turn, and existing OpenAI-compatible clients can connect just by changing base_url to http://localhost:4000/v1.