KOS Engine turns LLM into thin shell

This is the kind of anti-RAG project that actually has a thesis: if the graph is doing the reasoning, the model becomes a formatter. That makes the real question not which LLM you run, but whether the graph stays correct, auditable, and fast under messy real-world queries.

• –The traceable activation path is the strongest product value; provenance beats opaque vector search when teams need to debug why an answer won.
• –The repo’s quick-start still uses an OpenAI key for the thin LLM layer, so the fully local promise is still more modular design than default experience.
• –The launch’s 16/16 benchmark claim is encouraging, but the public README mostly shows a 10-point smoke test plus unification checks, so I’d treat it as prototype validation rather than broad proof.
• –The typo-recovery cascade and SymPy coprocessor are the practical bits that solve two pain points chat-only systems routinely botch: bad input and exact math.