Xin SI.

1. Use the dedicated link for Mac-to-NAS traffic and a separate interface for the office network. Do not turn the NAS into the office router.
2. Confirm 10Gbps negotiation at both ends, then copy a 10GB test file. Record write/read rates, checksums and the drive layout.
3. 10Gbps ÷ 8 = 1.25GB/s is a theoretical link ceiling, not file throughput. HDDs, SMB, adapters and small files all affect results.

Two 10GbE ports do not automatically give one task 20Gbps. Test a single link before considering aggregation.

1. Separate collections by company and project. Enforce access before retrieval; do not rely on the model to hide a company-wide index.
2. Keep model caches, databases and active indexes on the internal Mac SSD. Store originals and controlled backups on the NAS; link notes to their sources.
3. Two 8TB drives in RAID 1 provide about 8TB nominal capacity. RAID addresses a drive failure, snapshots some accidental changes, and separate/offline copies loss of the whole machine.
4. Restore three sample files and verify content and permissions. Test the UPS shutdown sequence too. A successful backup message is not a recovery test.

Company records remain under company control; the personal workstation is never the only copy.

1. Start with native Ollama and one quantized model around 8B–9B. Record the exact version, quantization, context length and runtime.
2. Test one user at 8K context first, then 16K and concurrency. Reserve memory for the OS, KV cache, document parsing and databases.
3. Use MLX as a comparison experiment, without loading two large models alongside Ollama. Compare quality, first-token latency, generation rate and peak memory on the same questions.

Fitting a model into memory does not prove task quality. No unmeasured tokens/s figures are published here.

1. Bind Open WebUI locally first and disable public registration. Use individual accounts, not the administrator account for daily chat.
2. Use a controlled private network and HTTPS for remote access. Approve devices individually and test revocation. Do not expose NAS administration or model ports publicly.
3. Email, contracts, payments and deletion produce drafts or proposed actions only, followed by human approval. Instructions inside external documents are not authorization.

The public discussion area is separate from the private workstation. Posting here does not connect to or read your NAS.

1. Input: the agenda, relevant files and previous decisions. Include only material needed for this meeting.
2. Output: three decisions, options and evidence for each, plus missing evidence. Retain file names and page numbers.
3. Acceptance: separate facts from recommendations; mark unanswered questions as unknown. Add decisions to the record only after owner confirmation.

Pilot target: record time saved against manual preparation and the number of corrections, not just generation speed.

1. Input: comparable periods, targets and product mappings. Keep tax-inclusive sales, net revenue and cash collections separate.
2. Output: value, volume, year-on-year change, target attainment and exceptions. Calculate in spreadsheets or code; use the model to suggest causes to investigate.
3. Acceptance: reconcile subtotals to the source, inspect three exceptions, and assign an owner, deadline and evidence to each action.

Do not let the model turn missing data into a plausible story.

1. Input: approved files from one project, with version, date, company, access level and source. Sample-check OCR and parsing first.
2. Output: an answer, cited excerpts, file names and page numbers; explicitly say when evidence is missing. Retrieval permissions must match file permissions.
3. Acceptance: include outdated versions, unanswerable questions and unauthorized requests in a 20–50-question test set. Record answer accuracy, citation correctness and access violations separately.

Suggested acceptance: zero unauthorized disclosures and traceable material numbers. Set accuracy targets for the task, not from someone else’s benchmark.