Project 3 — Annotated Bibliography (20+ Sources)

“Remember the days of old; consider the years of many generations; ask your father, and he will show you, your elders, and they will tell you.” — Deuteronomy 32:7 (ESV)

Chapter: 3 — Finding the Literature Weight: 10% of the course grade Due: End of Week 3 Submit: (1) a Zotero collection shared/exported, (2) references.bib (BibTeX export), (3) a formatted annotated bibliography (PDF or Markdown) in your venue’s citation style, and (4) your search log. Push all four to your research-portfolio Git repo and submit the link. See Appendix A for the submission flow and Appendix B for the Zotero workflow. Allowed tools: All discovery tools in the four-layer toolkit (FIND / UNDERSTAND / SNOWBALL / MANAGE). AI summarizers and chatbots are allowed as accelerators. AI policy: Every AI-suggested reference is tagged #unverified and must pass the §3.6 verification protocol before it counts. An unverified or fabricated citation in the submitted bibliography is research misconduct, not a point deduction — it can fail the project. Disclose any substantive AI assistance per Appendix C.

The Setup

You committed to a domain and a research question in Chapter 2. This week you find out whether anyone has stood where you’re standing — and you assemble the record of who did.

A working researcher does not start an experiment until they can describe the shape of the existing literature: the foundational papers everyone cites, the current state of the art, the open critiques, and the specific corner where their own contribution will sit. The annotated bibliography is how you build and prove that picture. It is also the raw material for next week’s literature review (Chapter 4) and the related-work section of the paper you began in Week 1. Do this well and Chapter 4 is half written. Do it badly and you’ll be searching from scratch in April.

The deliverable is at least 20 credible, verified sources for your chosen-domain applied-AI project, each captured in Zotero with a structured annotation, exported to BibTeX, and formatted in your venue’s style. “Credible” and “verified” are doing real work in that sentence — see the rubric.

Learning Targets

By completing this project, you will demonstrate that you can:

• Execute a systematic, multi-database literature search and log it reproducibly.
• Match a discovery tool to a goal across all four layers (FIND / UNDERSTAND / SNOWBALL / MANAGE).
• Snowball a citation graph in both directions to reach coverage keyword search can’t.
• Verify every reference against an authoritative index (dblp / Semantic Scholar / publisher) — and catch a fabricated AI citation before it reaches your bibliography.
• Write annotations that summarize, situate, and critique — not just describe.
• Maintain a clean Zotero library and export correct BibTeX for an acmart/IEEEtran paper.
• (Hard) Make the editorial judgment of what to exclude, and defend the boundary of your search.

Normal Tier

The Normal tier meets the deliverable bar for the week: a solid, verified, well-organized annotated bibliography.

Required features

• 20+ sources, all relevant to your research question, captured in a single Zotero collection.
• Source quality: a defensible mix — at least 3 foundational/seminal works, a body of recent (last ~3 years) state-of-the-art papers, and at least 1 critical or contrasting voice. At least one canonical applied-AI paper anchoring your domain (e.g., SWE-bench Verified for code review; an RL congestion-control paper for networking; a text-to-SQL/query-optimization paper for databases).
• Every reference verified per §3.6: exists in dblp or Semantic Scholar, metadata correct, and you’ve confirmed it says what you claim. Zero #unverified tags remain.
• An annotation per source (100–200 words) following code/annotation-template.txt: citation, summary in your own words, method/evidence (dataset/baseline/metric), relevance to your project, and a limitation/critique.
• A search log (code/search-strategy.txt): database, exact query string, date, hits, and hits-kept, for every search.
• Exports: a clean references.bib (BibTeX) and a formatted bibliography in ACM or IEEE style.

Normal-tier rubric (out of 100)

Medium Tier (+ up to 25% extra credit)

Stronger rigor and discovery craft.

• 30+ sources with documented two-direction snowballing: for at least 3 seed papers, show the backward (references) and forward (citers) expansion, with the tool used (Connected Papers / Research Rabbit / scite).
• Citation-context audit: run your 5 most important references through scite (or Semantic Scholar citation contexts) and note for each whether later work supports or contrasts it. Fold the contrasts into the relevant annotations.
• A coverage map: one paragraph + a small figure (a Connected Papers / Litmaps screenshot is fine) arguing that your set covers the field — naming the thinnest bucket and the query you’d run to fill it.
• The fabrication-drill artifact: include your Rep 6 table (5 AI references checked, with verdicts) as an appendix, with one paragraph on what it changed about your workflow.

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

Publication-grade — the work that makes this a real foundation for a submission. Hard always demands a judgment a tool can’t supply.

• A search-boundary memo (1–2 pages): state your inclusion and exclusion criteria explicitly and defend them. Which subfields did you deliberately exclude, and why? Where does your topic end? Which famous-but-irrelevant paper did you leave out on purpose, and what’s your justification? This is the editorial judgment that separates a curated bibliography from a pile of search results — and no tool can make these calls for you.
• A gap statement (1 paragraph) that emerges from the bibliography: read across your “relevance to my project” lines and state, in one precise sentence backed by cited absences (not a hunch), the gap your work will address. Name the nearest competitor and how you differ.
• Systematic-search reproducibility: your search log is complete enough that an independent researcher could rerun it and arrive at substantially the same set — exact query strings, databases, dates, and screening decisions.

Coach’s Note — The Hard tier is where you stop finding literature and start commanding it. Anyone with a browser can collect 30 papers. Telling me, with evidence, what you chose to leave out and why — that’s the move a thesis committee will ask you to make, and it’s the move an AI cannot make for you, because it requires owning a claim about the shape of an entire field.

Submission

Push to your research-portfolio Git repo and submit the link:

1. Zotero collection — shared link or exported .rdf/.ris.
2. references.bib — clean BibTeX export.
3. Annotated bibliography — formatted PDF or Markdown, ACM or IEEE style, one annotation per source.
4. Search log — your filled code/search-strategy.txt.
5. (Medium/Hard) coverage map, citation-context audit, fabrication-drill table, boundary memo, gap statement — as applicable.

Hints

• Start with FIND, not MANAGE. Get a candidate set first, then capture the keepers — don’t Zotero-clip every hit.
• Match the database to your field (§3.2). You’ll waste hours searching the wrong DL for your domain.
• Snowball early. The citation graph finds the paper that scoops you faster than any keyword. Do Rep 4 before you think you’re done.
• Verify in batches. When you’ve collected ~10, run all titles through dblp/Semantic Scholar in one sitting. Verification is boring; doing it in a block keeps you honest.
• Annotate as you read. A memory-written annotation drifts from the truth. Write it in the Notes field while the paper is open.
• Cap and curate. A well-read, well-annotated 22 beats an unread 80. Quality of engagement is graded, not the size of the pile.

What Mastery Looks Like

A mastered bibliography reads like the opening argument of a paper, not a list. Someone can scan your “relevance to my project” lines and see the gap forming. Every reference is real, correctly described, and earns its place. The search is reproducible. And — the tell of a real researcher — you can say, out loud, what you chose not to include and why. When your Chapter 4 reviewer asks “did you consider X?”, you have an answer with a date on it.

Coach’s Note — This is the most under-rated 10% in the course. Students rush it because “it’s just finding papers.” But the proposal you defend in Week 8 and the paper you present in Week 16 both stand on exactly the foundation you pour this week. Pour it level.

When You’re Done

• 20+ relevant sources captured in a Zotero collection.
• Every reference verified (exists / metadata / says what I claim) — zero #unverified tags.
• An annotation per source: summary + method + relevance + critique, in my own words.
• Coverage spans foundational / state-of-the-art / critical.
• Search log complete (database, query, date, hits).
• references.bib exported and imports cleanly into my Week-1 paper skeleton.
• Formatted bibliography in my venue’s style.
• Any AI assistance disclosed per Appendix C.

A theological footnote. Deuteronomy commands Israel to ask your elders, and they will tell you — to refuse the pride that imagines the present moment invented thinking. An annotated bibliography is that command in academic form: we remember the generations before us, and we report honestly what they told us. This is why verifying every citation is not bureaucracy but ethics. A fabricated reference does not merely break a rule; it bears false witness about what someone before you actually said. Honor them by getting it right — real authors, real findings, real limitations. The glory of kings, Proverbs says, is to search a matter out. The honesty of one is to report it as it truly is.