The paper proposes a criterion for determining which tasks are suitable for local LLM inference by introducing checkability as a task property. It presents Touchstone, a local-first pipeline that uses off-the-shelf SLMs and task-specific intrinsic checks to reject incorrect outputs.