Project 4 — Literature Review Draft

“This is the third time I am coming to you. Every charge must be established by the evidence of two or three witnesses.” — 2 Corinthians 13:1 (ESV)

Chapter: Chapter 4 · Due: end of Week 4 · Submit: to your Git repo + the manuscript, plus the matrix CSV (paths below) · Weight: 15% of the course grade (the largest deliverable before the proposal) · Allowed tools: Zotero, your literature databases, grounded AI extraction tools (Elicit / Consensus / scite / Semantic Scholar), Overleaf · AI policy: AI may extract and discover; you verify and own every cell and citation. Disclose AI assistance per Appendix C.

The Setup

You are six weeks from a proposal you will have to defend in front of peers, and twelve from an experiment you will have to run. Before you can claim your work is new, you have to prove — on the record, with cited evidence — that the field has a hole in it. That proof is the literature review. It is not a chore that precedes the real work; it is the real work’s foundation. Reviewers of your eventual paper will read this section first and decide, before they reach a single result, whether you understood the field well enough to add to it.

This week you turn last week’s annotated bibliography into a synthesized, thematic review of your chosen domain — and the comparison matrix that makes your gap undeniable. Do it well and your Related Work section is essentially written. Do it as a book report and you will rewrite it later under worse time pressure.

Learning Targets

You will demonstrate that you can:

• Synthesize a body of literature by theme — not summarize it paper by paper.
• Extract the load-bearing fields (problem / method / dataset / metric / limitation) from primary sources.
• Build a comparison matrix that makes the field’s structure — and its gap — legible at a glance.
• Identify and justify a research gap with cited absences, not assertions.
• Maintain citation integrity — every reference real, every cell verified, AI assistance disclosed.

Normal Tier

The deliverable bar for the week: a complete, honest literature review draft plus its matrix.

Required features

1. Comparison matrix (CSV, in the shape of code/comparison-matrix.csv) covering ≥ 12 papers — rows = papers; columns = problem, method, dataset/benchmark, metric & result, limitation, cite key. Every cell verified against the primary source.
2. Thematic review draft (~1,200–2,000 words) organized into 3–5 themes, each synthesizing several papers (no author-first topic sentences), drawing on ≥ 20 sources total.
3. A gap section (1–2 paragraphs) that names the unaddressed column / empty cell explicitly, backed by cited absences, and a one-sentence foothold for your planned work (CARS three moves).
4. A references.bib exported from Zotero, wired into your Overleaf acmart/IEEEtran manuscript so the reference list renders from keys. The no-admin default is the Zotero web library’s BibTeX export — no install, no admin (see Appendix A). If you’re on the local path and already have the Zotero desktop app, the Better BibTeX add-on is a nice optional convenience (auto-sync + stable cite keys) — but it is not required, and a plain web-library .bib earns full marks.
5. An AI-disclosure note stating which tools you used and for what (extraction vs. discovery vs. none), per the venue policy you’re targeting.

Normal-tier rubric (out of 100)

Medium Tier (+up to 25% extra credit)

Stronger rigor and breadth. Add any of:

• Expand the matrix to ≥ 20 rows and add a “baseline” column distinct from “metric & result,” so cross-paper comparability is explicit (and you flag the rows where the baseline is missing).
• A citation-context audit: run ≥ 5 key references through scite and report where the field’s actual treatment (supporting/contrasting/mentioning) diverges from how the papers are usually cited — and what that does to your gap.
• A coverage-gap statement: show the searches that came up empty (databases, queries, date ranges) so your “nobody has done X” is evidenced, not asserted.

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

Publication-ready judgment a tool cannot supply. Write a 1–2 page Synthesis Memo that does what no extraction engine can:

• Render a verdict on the field: not “here are the papers” but “here is what the body of testimony, weighed together, establishes — and where exactly it stops, and why that stopping point is the right place for new work.”
• Argue the comparability problem honestly: which results in your matrix are not comparable (different metrics, datasets, baselines) and how that uncertainty bounds the strength of your gap claim.
• Defend your gap against the strongest objection — the most plausible reason a reviewer would say “this has been done” — and answer it with cited evidence.

This memo is judgment, framing, and intellectual honesty about uncertainty. An AI can fill cells; only you can decide what the field means and stake your name on it.

Submission

Commit to your Git repo / portfolio:

• matrix/comparison-matrix.csv
• paper/related-work.tex (or Markdown) — the review draft, wired to references.bib
• references.bib (exported from Zotero — the web-library BibTeX export is the no-admin default; Better BibTeX is an optional convenience on the local path)
• notes/ai-disclosure.txt
• (Medium/Hard) the audit, coverage-gap statement, and/or synthesis memo

Submit the repo link + the rendered PDF of the review section.

Hints

• Build the matrix in the Comparison Matrix Builder widget first, export, then write from it. The writing goes fast once the grid is honest.
• The gap is almost always in the limitation column or an empty/absent column. Read down, not across.
• Auto-extract one column, then hand-verify — see Rep 5. Budget time for the verification; it always takes longer than the extraction.
• Rewrite every author-first topic sentence (Rep 9). Graders see the book-report failure mode instantly.

What Mastery Looks Like

A reader who knows nothing about your domain finishes your review and can state, in one sentence, exactly what hole you found and why it matters — and a reader who does know the domain nods, because every charge you laid was established by two or three witnesses.

Coach’s Note — The temptation under deadline is to inflate: cite more, claim a bigger gap, round a result up. Resist all three. A modest, true gap you can actually fill beats a grand one you invented. The reviewer rewards the honest, narrow claim. So does the One you’re really writing for.

When You’re Done

• Matrix ≥ 12 rows, every cell verified against the PDF
• Review is thematic — no author-first topic sentences
• Gap named explicitly, backed by cited absences
• Every citation confirmed real in dblp / Semantic Scholar / publisher
• references.bib from Zotero, wired into the manuscript, renders
• AI-disclosure note written and accurate
• Read the gap paragraph aloud — a stranger would know what you’ll do

A theological footnote. “Every charge must be established by the evidence of two or three witnesses” (2 Cor 13:1, ESV). A literature review is a courtroom, and you preside. The rule is not to count the witnesses or trust the loudest — it is to weigh them honestly together and render only the verdict the evidence supports. To cherry-pick the sources that flatter your thesis, or to let a fabricated citation stand, is not a methods error; it is to bear false witness against the readers who trust your review to walk the path behind you (Ex 20:16). The integrity you owe your sources is the integrity you owe your neighbor. And note what the hallucinating model can’t do: bear witness at all — because it cannot be held accountable for what it says. That accountability is yours. Carry it well.