The safety layer wrapped around an LLM app: input guards before the model, output guards after it, and least-privilege tools plus human review around the loop — so the model is powerful in the middle but never unsupervised at the edges.
LLM Guardrails & Safety — input guards · output guards · least-privilege tools · human-in-the-loop architecture by Blake Medulan · for more information contact blake@3halves-labs.com