Project P11 — Frontier vs Local Showdown

“O Timothy, guard the deposit entrusted to you.” — 1 Timothy 6:20 (ESV)

Chapter: 11 — Frontier vs Local: Cloud Power and Private AI Due: End of Week 11 Submit: A link to a shared folder or document (Google Drive, OneDrive, Dropbox, or a single PDF portfolio) containing your comparison spreadsheet, your screenshots, and your written report. No GitHub, no code — everything is a document, spreadsheet, or image. See Appendix A for setup and submission mechanics. Allowed tools: BoodleBox (sign in with your Concordia account at box.boodle.ai) as your default cloud venue — its multi-bot chat lets you @-mention a frontier model and a smaller/faster one in one conversation and compare them side by side; and a local model if you can run one (Ollama, LM Studio, Jan, or GPT4All — see Appendix A) for the maximum-privacy end. Fallback: any consumer AI you can reach from a browser or phone (ChatGPT, Claude, Gemini, Perplexity, Grok, DeepSeek), and a free-small-cloud model to stand in for local if you can’t install (see the fallback note below). Plus a spreadsheet app and a word processor. AI policy: You will use AI heavily and evaluate it honestly. Every deliverable that touches AI carries a one-line AI usage note. Hard rule: no real confidential or regulated data goes into any cloud tool — use fictional stand-ins throughout. You choose the venue; you own the verdict.

The Setup

You are Dana Ruiz, the HR coordinator at Cedar Ridge Manufacturing, a 300-person company with one HR department: you and your director. Your days are a flood of writing — job postings, policy summaries, onboarding packets, benefits explainers, the occasional delicate rewrite of a manager’s too-blunt performance note. AI could clearly save you hours a week. You’ve been told to “figure out how we should use it.”

But here’s your knot. Some of your writing is harmless — a job posting, a summary of a public benefits guide. And some of it is the most sensitive data in the whole company: employees’ Social Security numbers, medical-leave letters that name a diagnosis, performance reviews with people’s names on them, salary bands. You already know, in your gut, that pasting a medical-leave letter into a free chatbot would be wrong — maybe illegal. But you also don’t want to ban a genuinely useful tool over a fear you haven’t actually tested.

So you’re going to run a showdown. You’ll take a realistic slice of your weekly work, run each task on a frontier cloud model and a local model (private, on your own machine), and measure what you actually get from each — quality, speed, privacy. At the end you’ll write your director a straight recommendation: here’s what goes to the cloud, here’s what stays local, and here’s why. You are the steward of every employee’s most private information. You will not hand the keys to a tool you haven’t tested.

Cloud-only fallback (no admin rights / locked-down laptop): if you can’t install a local app, run each task on a frontier cloud model and a free-small-cloud model (a “mini/nano/flash” tier, or free DeepSeek/Qwen). In BoodleBox you can run both in one multi-bot chat: type @ to add a frontier model and @ again to add a smaller/faster one, so a single prompt returns both answers to compare. The small model stands in for “local” for the capability and speed comparison. For the privacy comparison, you’ll reason about where each task’s data would go — and mark the distinction this week turns on: BoodleBox is a vetted cloud (FERPA/SOC-2, doesn’t train on your data, as of 2026), a very different posture from a consumer free tier, but still not on-device local (both venues here are still cloud). (No BoodleBox? Run a frontier free tier and a free-small model in two browser tabs.) This is a fully acceptable path — say so up front in your report.

Setup (the starter)

This chapter ships two starter artifacts in code/. Copy them into your project folder and make them yours:

• code/cloud-or-local.csv — the routing worksheet. One row per task: sensitivity tier, volume, whether it needs the newest capability, offline need, the recommended venue, and why. Six example rows are filled in; the rest are yours.
• code/data-sensitivity-guide.txt — the four-tier “Can this go to the cloud?” guide (Public / Internal / Confidential / Regulated) with the one-line test. This is your rulebook for classifying every task.

You will also invent your own task inputs — all fictional. Do not use a single real employee’s data.

Learning Targets

You will demonstrate that you can:

• Run the same task in two venues (frontier cloud and local / free-small stand-in) and produce evidence — screenshots and saved outputs — of the difference.
• Score the tradeoff honestly across three axes: quality, speed, and privacy.
• Classify data sensitivity correctly and route each task to a venue it is allowed to use — never sending confidential or regulated data to a consumer cloud tool.
• Turn the results into a defensible recommendation — when to reach for the frontier, when to stay local — that a real manager could adopt.

Normal Tier

Goal: Run a five-task showdown, capture the evidence, complete the routing worksheet, and write a clear when-to-use-each recommendation.

Required features

1. Pick five tasks from Dana’s world that span at least three sensitivity tiers — e.g., a public job posting (Public), a summary of a published benefits memo (Public/Internal), an internal onboarding checklist (Internal), a rewrite of a named, fictional performance note (Confidential), and a leave letter that references a fictional condition (Regulated). Write each task’s fictional input yourself.
2. Run each task in two venues — a frontier cloud model and a local model (or the free-small stand-in). In BoodleBox you can run both cloud models in one multi-bot chat (type @ to add each), which makes the side-by-side screenshot easy. Capture a screenshot of each run and save every output. For the Confidential/Regulated tasks, you must not run them in a consumer cloud tool; run them locally, or in the vetted institutional tool — at Concordia that’s BoodleBox (FERPA/SOC-2, does not train on your data, as of 2026), using your fictional input — or (fallback) run a sanitized/fictional version and note where the real data would go.
3. Comparison table (spreadsheet or doc) scoring each task, each venue, on quality (1–5), speed (fast/ok/slow), and privacy (data left the building / stayed).
4. Completed cloud-or-local.csv — fill in a row for each of your five tasks with the recommended venue and a one-line why.
5. Recommendation — a written conclusion (roughly 400–600 words) answering: for Cedar Ridge, which kinds of HR work should go to the frontier cloud, which must stay local, and why? Ground it in your evidence, not vibes.
6. AI usage note — one honest line per task: what you asked, which venue, what you verified. Include at least one place where you spot-checked an AI output and caught (or ruled out) an error or made-up detail.

Normal-tier rubric (out of 100)

Medium Tier (+up to 25% extra credit)

Make the comparison sharper.

1. Add volume and cost. For each task, estimate how many times a month Dana does it, and reason about the cost: at consumer-plan prices (a ~$20/month tier as of mid-2026 — cite the vendor’s current page, don’t trust a printed number), which tasks are cheap on the cloud and which would be better local (free per use)? Add a “monthly volume” and “cost note” column.
2. A second local model. Run two of your tasks on a second local (or free-small) model of a different size and compare. Does a bigger local model close the gap with the frontier? Note the quality-vs-size tradeoff as a mid-2026 snapshot.
3. The offline test. Turn off your Wi-Fi and document which of Dana’s tasks still work. Add a short paragraph: which parts of her workflow need an offline-capable fallback?

Hard Tier (+up to 25% additional extra credit)

This is the judgment no model can produce for you, and it is graded as such.

Write venue-routing-memo.pdf (or .docx), one page, addressed to “the HR Director, Cedar Ridge Manufacturing.” Make and defend a single, concrete policy recommendation: which categories of HR work are permitted on the frontier cloud, which must stay local or on a vetted enterprise tool, and which must never touch AI at all. (If you’re at Concordia, your real-world vetted enterprise tool is BoodleBox — FERPA/SOC-2, doesn’t train on your data as of 2026; you may cite a tool like it as Cedar Ridge’s vetted middle option.) You must:

• Name at least one category for each of the three buckets (cloud-OK / local-only / never), and give the specific reason tied to a sensitivity tier and a law or duty (PII, HIPAA-style health info, FERPA-style records, or plain confidentiality).
• State the failure mode you are guarding against, and name who at Cedar Ridge would answer for it if it went wrong.
• Tie the recommendation explicitly to the spine rule — you choose the venue; you own the verdict — and to at least one concrete control (the sensitivity guide, a “never paste” list, a designated vetted tool — a BoodleBox-style FERPA/SOC-2 platform — for confidential work, a required-approval step).

A memo that says “use the cloud for everything” or “ban AI entirely” fails. The grade is in the discrimination — showing you can tell a cloud-OK task from a local-only one and defend the line to a boss who will be held responsible for it.

Submission

Put everything in one shared folder (Google Drive / OneDrive / Dropbox) or one combined PDF, and submit the link per Appendix A:

project-11-frontier-vs-local/
  comparison-table.xlsx        (or a table inside the report)
  cloud-or-local.csv           (your completed worksheet)
  recommendation.docx/.pdf      (the Normal-tier write-up)
  screenshots/                  (one per task per venue, clearly labeled)
  venue-routing-memo.pdf        (Hard tier only)

Everything is a document, spreadsheet, or image — no code, no GitHub. Use fictional data only; never include a real person’s information in a screenshot or file.

Hints (Read Before You Begin)

• Classify first, run second. Decide each task’s tier before you touch a tool. If you run first, you’ll be tempted to paste sensitive text into the convenient cloud “just to compare” — which is exactly the mistake this project exists to break.
• Screenshots are your evidence. A claim like “the local model was slower” is worth little; a labeled screenshot with a timestamp is worth a lot. Capture as you go — you won’t want to re-run everything later.
• Sanitize your Confidential/Regulated inputs. Write them as obviously fictional (fake names, fake conditions, fake numbers). The point is to practice the routing, not to create real risk.
• “Free” is not “private.” If your fallback uses a free cloud model, say so, and note in your report that its privacy is simulated, not real. Honesty about the limits of your setup is part of the grade.
• Let the evidence pick the winner. Sometimes the local model will surprise you and be plenty good. Sometimes the frontier will clearly win. Report what you actually found, not what you expected.

What Mastery Looks Like (Beyond the Rubric)

A mastered submission is one where I can read your recommendation and your memo and tell that you are the one deciding — that you caught the moment the convenient cloud tool would have taken an employee’s diagnosis, and you said no; that you can name which HR tasks belong on the frontier and which must stay on Dana’s own machine, and defend the line to a director who will answer for it. The screenshots and the table are table stakes. The routing judgment — sensitivity first, capability second — is the project.

Coach’s Note — The tempting shortcut this week is to make local look either useless (so you can justify pasting everything into the cloud) or magical (so you never have to touch the cloud). Resist both. The honest finding is that each door is right for different work — and a steward’s whole skill is knowing which is which. Report the real tradeoff. Your credibility for the rest of the course rides on whether I believe your Week-11 numbers.

When You’re Done (a short checklist)

• Five tasks, ≥3 sensitivity tiers, each run in both venues.
• Screenshots + saved outputs for every run, clearly labeled.
• No confidential/regulated data sent to a consumer cloud tool (fictional data throughout).
• Comparison table scores quality, speed, and privacy.
• cloud-or-local.csv completed with a reason per task.
• Recommendation answers “when to use each,” grounded in evidence.
• AI-usage notes are honest; at least one verification documented.
• (Hard) The memo defends a real cloud/local/never line and names who’s accountable.

A theological footnote. Paul’s charge to Timothy is to “guard the deposit entrusted to you” (1 Timothy 6:20, ESV) — a deposit being something valuable that belongs to another and is placed in your keeping. Every file in this project stands for a real deposit: an employee’s diagnosis, a colleague’s review, a person’s Social Security number — none of it yours, all of it entrusted to Dana, and to you. The frontier model is a good gift, and part of faithfulness is knowing which gifts a given task is allowed to accept. To route the public work to the cloud and keep the private work on your own machine is not fussy compliance; it is the Eighth Commandment kept with a keyboard — guarding your neighbor’s secret because it was never yours to spend. Choose the venue. Own the verdict. Give the deposit back intact.

See you next week.