Complete procedures with a way back.

Build a local-first analytics assistant that discovers only curated PostgreSQL views, asks OpenAI for a structured SELECT proposal, validates a conservative SQL subset, requires a human to approve the exact query, enforces a restricted database role and read-only transaction, limits time and rows, and emits local table or chart JSON.

Create a reproducible Linux and NVIDIA benchmark that separates time to first meaningful content, end-to-end latency, exact server-reported token throughput, memory phases, deterministic task quality, long-context retrieval, structured output, and tool-call correctness. Results remain tied to immutable artifacts, scorer revisions, deployment flags, and host state.

Build a bounded, repeatable compatibility suite for Ollama 0.32.0, LM Studio 0.4.1, and vLLM 0.23.0. The guide tests advertised models, non-streaming and streaming chat, strict structured output, forced tool calls, and the Responses endpoint while keeping provider documentation, locally measured evidence, and product positioning visibly separate.