Chapter 3 — Reps

This week’s reps are not warm-up drills you throw away. Every one of them builds a piece of the Annotated Bibliography you’ll submit on Friday. By the end you’ll have a search log, a Zotero library, and 20+ verified, annotated sources for your domain’s applied-AI project. The keyboard is the gym; the literature is the weight.

Ground rules.

• All reps target your own research question and domain from Chapter 2. No generic practice — this is your real project moving forward.
• AI is allowed this week as an accelerator, with one absolute rule: every reference an AI suggests is tagged #unverified and must pass the §3.6 verification protocol (exists / metadata correct / says what was claimed) before it enters your bibliography. An unverified AI citation in your deliverable is treated as a fabrication.
• Keep a search log from Rep 1. Database, exact query, date, hits kept. Template: code/search-strategy.txt.
• Capture everything in Zotero as you go (Appendix B). A paper you read but didn’t capture is a paper you’ll re-find at 1 a.m.

Rep 1 — Build the keyword grid and Boolean query

Open the Literature Search Builder widget on the chapter page. Pull the 2–4 core concepts out of your research question and, for each, list every synonym you can think of — including the terms adjacent subfields use.

Concept A (method): ______ OR ______ OR ______
Concept B (task):   ______ OR ______ OR ______
Concept C (domain): ______ OR ______ OR ______

Assembled:  (A1 OR A2 OR A3) AND (B1 OR B2) AND (C1 OR C2)

Write-up: Paste your final Boolean query string. In one sentence, name the one synonym you almost forgot — the one a different subfield uses for your concept.

Rep 2 — Search at least three databases, and log every query

Run your query (adapted to each system’s syntax) in at least three FIND-layer tools chosen for your field: Google Scholar plus two of {ACM DL, IEEE Xplore, arXiv, dblp, OpenReview}. Skim titles/abstracts; keep the ones worth a real read.

| Date | Database | Exact query | Hits | Kept |
|------|----------|-------------|------|------|
| ...  | ACM DL   | ...         | 312  | 9    |

Write-up: Which database returned the most relevant hits for your topic, and why does that make sense for your field? (HCI/SE → ACM DL; networking/architecture → IEEE Xplore; ML → arXiv + OpenReview.)

Rep 3 — Triage with an UNDERSTAND tool

Take your kept set and run it through Semantic Scholar (TLDR summaries) or Elicit (extract method/dataset/result columns) or Consensus. Use the summaries to decide which papers earn a full read.

Write-up: Pick one paper. Compare the tool’s TLDR/summary to what the abstract actually says. Did the tool get it right? Note one thing it dropped or distorted. (This is the §3.5 discipline: summary is a lead, not a citation.)

Rep 4 — Snowball a seed in two tools and compare

Choose your single best seed paper. Run backward (its references) and forward (its citers) snowballing in Connected Papers, then do the same seed in Research Rabbit. The neighborhoods will differ.

Write-up: Name two papers that appeared in one tool but not the other. Which would you have missed with keyword search alone? The union of the two neighborhoods is your expanded candidate set — add the keepers to Zotero.

Rep 5 — Audit how your key papers were cited

Run three of your most important references through scite (or read the “Cited by” context on Semantic Scholar). Note scite’s classification — supporting, contrasting, or mentioning.

Write-up: Did any paper you assumed was foundational get contrasted (challenged) by later work? That contrast is a live gap — record it; it may be your opening in Chapter 4.

Rep 6 — The fabrication drill (do this one slowly)

Ask a general-purpose chatbot for five references on a niche corner of your topic. Then verify each one with the §3.6 protocol: search the exact title in dblp and Semantic Scholar.

| # | Title as given | Exists? | Metadata correct? | Verdict |
|---|----------------|---------|-------------------|---------|
| 1 | ...            | yes/no  | yes/no            | real / fabricated / errored |

Write-up: How many of the five were entirely fabricated? How many real-but-errored? Compare your count to the chapter’s figures (55%/18% fabricated; 87% with metadata errors). One sentence on what this changes about how you’ll use AI for the rest of the semester.

Rep 7 — Set up Zotero and capture ten sources

Install Zotero + the browser Connector if you haven’t (Appendix B). Create a collection for your project. Capture at least ten verified keepers — metadata + PDF — and tag each with its sub-topic. Tag anything AI-suggested #unverified until it passes Rep 6’s protocol.

Write-up: Export the collection to BibTeX (references.bib) and confirm it imports cleanly into the paper skeleton you started in Week 1. Paste two BibTeX entries. (Reference: code/references.bib.)

Rep 8 — Write three real annotations

Pick three captured papers and write a full structured annotation for each in the Zotero Notes field, following code/annotation-template.txt: citation, summary in your own words, method/evidence, relevance to your project, limitation/critique. 100–200 words each.

Write-up: For each, the single most important line — relevance to your project — in one sentence. Is it your baseline, a method to adapt, or the gap you’re filling?

Rep 9 — Find the closest competitor and the oldest ancestor

From your collection, identify (a) the most recent paper closest to your exact idea — your nearest competitor — and (b) the earliest seminal work your line of research descends from.

Write-up: Two sentences. How does your planned contribution differ from the nearest competitor? And what did the seminal paper establish that the whole field still rests on?

Rep 10 — Coverage check and dedupe

Pull all sources into one list. De-duplicate (the same paper appears as arXiv preprint and published version — keep the published one, note the arXiv id). Then sanity-check coverage: do you have the foundational papers, the recent state of the art, and at least one critical/contrasting voice?

Write-up: Which of the three coverage buckets (foundational / state-of-the-art / critical) is thinnest, and what’s your next query to fill it? Tools like Litmaps can set an alert so new work in the gap finds you.

Done? One Last Thing.

Assemble the full Annotated Bibliography. This is the capstone rep and it rehearses Friday’s deliverable directly. Bring your Zotero collection to 20+ verified sources, each with a complete annotation in the Notes field. Then:

1. Confirm every reference passed the §3.6 verification protocol — zero #unverified tags remain.
2. Confirm coverage across foundational / state-of-the-art / critical (Rep 10).
3. Export references.bib (BibTeX) and a formatted bibliography (your venue’s style — ACM or IEEE).
4. Re-read your “relevance to my project” lines end to end. Do they, read together, start to argue for a gap? They should. That argument is the seed of your Chapter 4 literature review.

If you can do that, you’ve not just done the reps — you’ve built the foundation the next thirteen weeks stand on. That’s the deliverable. That’s the point.

Up next: Project 3 — the Annotated Bibliography (20+ sources), 10% of your grade.