Chapter 4 — Reps

Every rep this week moves your own literature review forward. By the time you finish the capstone rep, this week’s deliverable is assembled — not started. The keyboard is the gym; the matrix is the bench press. Reps in order.

Ground rules

• Work on YOUR domain and YOUR sources — the 20+ you collected in Chapter 3, tracked in your Zotero library.
• Verify every cell and every citation against the primary source. Auto-extraction is a first draft, never a fact. Confirm existence + metadata in dblp / Semantic Scholar / the publisher.
• Write a short reflection after each rep — two or three sentences — into your research portfolio (a notes/ file in your Git repo). The reflection is graded as part of the process, not the answer.
• AI policy: grounded tools (Elicit, Consensus, scite) are allowed for extraction and discovery; bare-LLM-suggested citations are leads, not sources. Disclose AI assistance per Appendix C.

Rep 1 — Pass-1 triage your whole set

Run the five-minute bird’s-eye pass on all 20+ of your sources. For each, record the five C’s (Category, Context, Correctness, Contributions, Clarity) in one line, then a triage verdict: keystone / Pass-2 / cite-only / drop.

# notes/triage.txt  (one row per paper)
citekey | category | contribution (1 line) | verdict

Reflection: Which papers surprised you by being weaker (or stronger) than the title promised? How many survived to Pass 2?

Rep 2 — Identify your 3–5 keystones

From the survivors, pick the 3–5 papers your own work will directly stand on or compete with. Justify each in one sentence: why is this a keystone and not just a Pass-2 read?

Reflection: A keystone is usually one you’ll baseline against or whose limitation becomes your gap. Did naming keystones change what you think your gap is?

Rep 3 — Fill five extraction sheets

Copy code/extraction-sheet.txt once per keystone and fill all five fields (problem, method, dataset/benchmark, metric & result, limitation) — in the paper’s own terms and in yours.

for k in keystone1 keystone2 keystone3 keystone4 keystone5; do
  cp code/extraction-sheet.txt "notes/extract-$k.txt"
done

Reflection: Which field was hardest to fill? (For most students it’s the baseline hiding inside “metric & result.” If you can’t name the baseline, the result isn’t comparable.)

Rep 4 — Build the comparison matrix in the widget

Open the Comparison Matrix Builder below the chapter. Enter your keystones (and as many Pass-2 papers as you can). Fill the five columns. Toggle “highlight gaps” and read down the columns. Export to CSV in the shape of code/comparison-matrix.csv and commit it to your portfolio.

Reflection: Which column or cell stayed conspicuously empty across many rows? Write the one sentence that turns that emptiness into a research question.

Rep 5 — AI-extraction-vs-verification drill

Pick one column (say, “dataset/benchmark”). Auto-fill it for ~5 papers with Elicit (or Consensus Study Snapshots). Then hand-verify each cell against the PDF. Log every discrepancy.

# notes/extraction-audit.txt
citekey | tool said | PDF says | correct? | note

Reflection: How many cells were wrong or over-claimed? Tie your count to the Walters & Wilder (2023) figures (55% GPT-3.5 / 18% GPT-4 fabricated). Grounded tools err less — by how much, in your sample?

Rep 6 — Citation-context audit with scite

Run three of your keystone references through scite Smart Citations. Reclassify how the field actually treats each: supporting / contrasting / mentioning.

Reflection: Did any paper you assumed was “the consensus” turn out to be contrasted by later citers? What does that do to your gap?

Rep 7 — Verify every citation against a real index

Take your full reference list and confirm each entry exists with correct metadata in dblp or Semantic Scholar. Flag anything an LLM ever suggested and re-confirm it independently.

# notes/citation-verify.txt
citekey | found in dblp? | metadata correct? | source URL

Reflection: Any phantom or wrong-metadata entries? Remember: ICCV rejects papers with non-existent citations without review. Better to catch it here.

Rep 8 — Cluster the matrix into themes

Group your rows into 3–5 themes by what they have in common (shared benchmark, shared method family, shared metric). Name each theme. These become the subsections of your review body.

Reflection: Does any single paper resist every theme? A loner row is often either the most novel work in the field — or a clue your gap is real.

Rep 9 — Rewrite three author-first topic sentences

Find three paragraphs (or planned paragraphs) that open with an author’s name. Rewrite each to open with an idea about the field, then marshal the witnesses to support it.

Before: "Smith (2025) trained an agent on SWE-bench..."
After:  "Existing agent benchmarks reward task completion but
         ignore review quality [Smith 2025; ...], leaving open..."

Reflection: This single edit is the difference between a B and an A on this deliverable. How many of your planned paragraphs are still author-first?

Rep 10 — Generate references.bib from Zotero

Export your collection to references.bib, drop it next to your Overleaf manuscript, and cite two keystones by key. Confirm the reference list renders. Use code/references.bib as a shape reference. No-admin default: export BibTeX straight from the Zotero web library (collection → ⋯ More → Export → BibTeX) — no install needed. If you’re on the desktop app (local path), the Better BibTeX add-on auto-syncs the file and pins stable keys, but it’s an optional convenience, not a requirement. See Appendix A.

Reflection: Did any cite keys churn or collide? On the no-admin path, the fix is to pick a sensible key once and leave the item’s author/year alone so it stays put. Stable keys are why Zotero stays your single source of truth — confirm you never hand-edited the .bib.

Done? One Last Thing.

Assemble the gap paragraph. Pull the empty cell / unaddressed column from your matrix, back it with the cited absences you found in Reps 6–8, and write the two-paragraph passage that (a) names what the field collectively knows, and (b) names the specific hole your work will fill — with a one-sentence foothold for how you’ll occupy it (Swales CARS Move 3). This paragraph is the spine of Project 4; everything else in the review points at it.

Read it aloud. If a stranger could finish it and not know exactly what you’re going to do something about, it isn’t done yet.

Up next: Project 4 — the Literature Review Draft (15%).