Design a privacy-first local chat setup with quantization guidance

You are a senior engineer who sets up privacy-preserving local AI. Build a privacy-first local chat configuration and quantization guide.

My context:
- Sensitivity: [WHY IT IS PRIVATE — e.g. 'client NDA docs', 'personal health notes', 'unreleased financials']
- Hardware: [GPU and VRAM / CPU only / Apple Silicon unified memory]
- Desired capability: [GENERAL CHAT / SUMMARIZATION / CODE — be specific]
- Quality vs speed: [BEST I CAN RUN / BALANCED / FASTEST]
- OS: [macOS / Linux / Windows]

Produce:
1. A quantization primer in plain terms — what q4_K_M, q5, q8, and f16 mean for quality, size, and speed; how to choose based on my VRAM/RAM. No hand-waving — show the size math for a representative model.
2. A concrete recommendation: model and tag for my hardware and capability, with the exact ollama pull command and approx disk/RAM cost. Offer a heavier fallback if I have headroom and a lighter one if I am tight.
3. A tuned run configuration: a system prompt suited to my capability, plus context length, temperature, repeat-penalty, and keep-alive values — each with a one-line reason.
4. A privacy configuration that locks the setup to local-only:
- Run Ollama bound to 127.0.0.1 (set OLLAMA_HOST to 127.0.0.1:11434), not 0.0.0.0.
- Disable any telemetry; confirm honestly what data the model layer could send (a local model sends nothing to a server, but note if the installer or UI does).
- Where transcripts and model files live on disk, and how to purge them.
5. A privacy audit checklist — concrete commands or checks to verify nothing is leaving the machine (localhost binding, no unexpected outbound connections, no cloud sync of the model directory).

Rules:
- Privacy claims must be verifiable, not marketing. If a setting's privacy behavior is uncertain, say 'verify' rather than asserting it.
- Do not overclaim model quality. Be honest that a local quantized model is not a frontier cloud model.
- Recommend only tags that plausibly exist; flag 'verify the tag exists'.

Output: the quantization primer, the model recommendation, the run config, the privacy configuration, and the audit checklist.

Success signal: the output is good only if the quantization guidance includes real size math, the setup is locked to localhost with verifiable checks, and every privacy claim is either verifiable or explicitly flagged 'verify'.