Jev & balo_engine — The deterministic AI nervous system.

Part of the BalonixOS core: language models interpret meaning while deterministic Rust enforces schemas, safety gates, and three-beat state transitions.

Integrated OS modules

system/llm-agent/
├── orchestrator/   # Rust (Jev Harness & safety gate)
├── balo_engine/    # OS talents, router, EmbeddingGemma-2
└── runtime.nix     # Declarative llama-server on-demand unit
01

8-Layer Rust Harness

02

0 MB VRAM idle

03

Three-beat transactions

Typed OS talents instead of brittle shell hacks

Every capability is modeled as a structured talent checked by a post-LLM safety gate before execution.

Three beats from live state to NixOS

Instant runtime feedback, session state, and declarative Nix modules share a single source of truth.

On-demand local inference runtime

The managed llama-server runtime spins up for prompts or vector searches and frees GPU memory when idle.

Can an LLM invent arbitrary shell commands?

No. Jev restricts execution to registered typed talents guarded by explicit parameter checks.

What role does EmbeddingGemma-2 play?

Fast local vector indexing and resilient semantic matching across languages and typos.

Is Jev distributed separately from BalonixOS?

No — Jev and balo_engine live directly inside BalonixOS under GPL-3.0.