Project P7 — Summarize and Verify

“they received the word with all eagerness, examining the Scriptures daily to see if these things were so.” — Acts 17:11 (ESV)

Chapter: 7 — Everyday Productivity II: Research, Documents, and Data Due: End of Week 7 Submit: A link to a shared folder (Google Drive / OneDrive / Dropbox) or a single combined PDF. Non-technical — no GitHub, no code. See Appendix A. Allowed tools: Any general assistant with document upload (ChatGPT, Claude, Gemini) and, optionally, a spreadsheet tool. (Tool reference: Appendix B.) AI policy (Phase 1): AI summarizes; you check every claim against the source before you trust it. The provided document is fictional and safe. Do not upload real confidential documents to a consumer tool — read Appendix C.

The Setup

Meet Marcus Bell, an operations manager at the Meridian Group. The company’s 2025 Hybrid Work & Productivity Report just landed in his inbox, and he has to brief the leadership team on it in two days — accurately, because leadership will set hybrid policy and decide whether to commission a real-estate consolidation study based on what he tells them. He does not have two days. He has about ninety minutes.

This is the perfect job for AI: turn a dense, number-heavy report into a brief, fast. It is also the perfect trap, because a summary that sounds authoritative and gets one number wrong could send real money in the wrong direction. This week you learn the single most important research habit in the whole course — the Berean move: receive the summary eagerly, then examine it against the source to see if these things are so. A summary you haven’t verified is a rumor with good formatting.

Setup (the starter)

This chapter ships four starter files in code/:

• code/sample-report.txt — the fictional Meridian Group hybrid-work and productivity report. Your source of truth.
• code/quarterly-figures.csv — the numbers behind the report, for checking any figure the AI cites.
• code/ai-summary-flawed.txt — a pre-made AI summary of the report with several errors seeded in. You’ll audit it.
• code/grounding-checklist.txt — the step-by-step method for grounding a summary claim-by-claim in the source.

Copy all four into your project folder.

Learning Targets

You will demonstrate that you can:

• Summarize a dense report with AI into a brief a busy decision-maker can act on.
• Ground a summary — check each claim against the source and mark it supported, unsupported, or wrong.
• Audit someone else’s summary — find the seeded errors in the provided flawed summary, including a fabricated statistic and a misread figure.
• Verify numbers — cross-check any cited figure against the underlying data.
• Own the verdict — deliver a brief you’d stake your name on to a leadership team making a real decision.

Normal Tier

Goal: Produce your own AI summary of the report, verify it claim-by-claim against the source, and separately audit the provided flawed summary.

Required features

Produce a folder named p7-summarize-verify containing:

1. my-summary.docx. Upload code/sample-report.txt to an assistant and generate a one-page leadership brief (key findings, the recommendations and the money at stake, the risks). Keep the raw AI version and your final edited version.
2. grounding-audit.xlsx (or a table). Using code/grounding-checklist.txt, take each substantive claim in your summary and mark it Supported (quote the source line), Unsupported (not in the source), or Wrong (contradicts the source). Every “Supported” needs the actual source quote.
3. The flawed-summary audit — flawed-audit.docx. Read code/ai-summary-flawed.txt against the report and the figures. Find every error: a fabricated statistic, a misread number, an overreach the source doesn’t support, and a claim stated with false confidence. For each, name the error type and the correct fact.
4. Number checks. For at least three figures cited anywhere, verify them against code/quarterly-figures.csv and show your check.
5. brief-final.docx + AI usage note. Your verified, leadership-ready one-pager, plus one honest line: what you asked, what the AI got wrong, what you verified.

Normal-tier rubric (out of 100)

Medium Tier (+up to 25% extra credit)

Pick one or more:

1. Force the citations. Rewrite your summary prompt to require the AI to cite the section or line for each claim. Does grounding get easier? Does the AI still invent a citation? (Fabricated citations are a classic failure — catch one.)
2. Two summaries, one document. Summarize the report with two different tools and compare which hallucinated more. Note the difference.
3. Ask the data a question. Use a spreadsheet AI feature (or a prompt over the CSV) to answer a question the narrative report doesn’t directly state, then verify the answer against the numbers yourself.

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

This is the judgment the machine cannot make for you.

Write board-note.docx — a short, honest note to the leadership team that accompanies your brief and states what you are and are not confident about, and why. You must:

• Separate fact from inference. Which statements are directly in the report, and which are the AI’s (or your) interpretation? Label them.
• Flag the decision-critical numbers and say how you verified each — because leadership is about to make a real-estate and hybrid-policy decision on them.
• Name one thing you would not let AI decide here and hand to a human, tying it to the spine rule.

Leadership doesn’t need a confident summary; it needs a trustworthy one. The grade is in your calibration — knowing, and admitting, exactly how far your verification reaches.

Submission

One shared folder or a combined PDF; submit the link per Appendix A:

p7-summarize-verify/
  my-summary.docx        — raw + edited
  grounding-audit.xlsx   (or a table)
  flawed-audit.docx      — the seeded-error audit
  brief-final.docx       — the leadership-ready one-pager
  board-note.docx        (Hard tier)

Do not upload your own employer’s documents, HR data, or anything confidential — the fictional Meridian report is here for exactly that reason. If in doubt, Appendix C.

Hints (Read Before You Begin)

• A summary is a claim, not a fact. Treat every sentence the AI produced as something to check, not something to trust because it reads well.
• Numbers are where it bites. Models paraphrase numbers wrong constantly. Any figure headed for a decision gets checked against the source data, no exceptions.
• “Supported” means you can quote it. If you can’t point to the line, it’s not supported — it’s plausible, which is different.
• The flawed summary is calibrated. Some errors are obvious; at least one is subtle and stated with total confidence. That subtle one is the real skill.

What Mastery Looks Like (Beyond the Rubric)

A mastered submission is one where your final brief is trustworthy — every claim traceable, every decision-critical number checked, every AI overreach caught and corrected. You found all the seeded errors, including the confident-but-wrong one most people skim past. And your Hard-tier note is honestly calibrated: it tells leadership what’s solid, what’s inference, and where a human still has to decide. Anyone can get a summary in ten seconds. You can get a true one.

Coach’s Note — The reason this is the most important habit in the course is that it’s the one you’ll be most tempted to skip. The summary looks done. It reads beautifully. Checking it feels like distrust. But the Berean habit isn’t cynicism — it’s respect for the people relying on you. Receive the draft eagerly. Then examine it to see if it’s so.

When You’re Done (a short checklist)

• AI summary produced; raw and edited versions kept.
• Every substantive claim grounded (Supported with a quote / Unsupported / Wrong).
• All seeded errors in the flawed summary found, typed, and corrected.
• Three+ figures verified against the data, checks shown.
• Final brief accurate and leadership-ready.
• Honest AI usage note; nothing confidential uploaded.
• (Hard) board-note.docx separates fact from inference and calibrates confidence.

A theological footnote. Luke praises the Bereans as “more noble” for a specific reason: when Paul — an apostle — preached to them, they didn’t just accept it because it sounded authoritative. They examined the Scriptures daily to see if these things were so. Notice the posture: eager reception and diligent checking, held together. That is exactly the stance a professional takes toward an AI summary. Not suspicion that refuses the tool, and not credulity that swallows it whole — but the noble middle: receive it gladly, then verify it against the source, because the people downstream of your brief are trusting you to have looked. Leadership will act on what you tell them. Be found to have examined it.

See you next week.