Chapter 14 — Writing the Research Paper

“The problem is, why is it a problem, the startling sentence, the implication.” — Kent Beck, on the four-sentence abstract (OOPSLA 1993)

“And the LORD answered me: ‘Write the vision; make it plain on tablets, so he may run who reads it.’” — Habakkuk 2:2 (ESV)

Why This Matters

You have done the work. You found a gap, framed a falsifiable question, reviewed the literature, pinned an environment, ran a pilot, executed the full experiment, logged every number with provenance, ran honest statistics, and interpreted what you found without overclaiming. That is thirteen weeks of real research sitting in a Git repository and a pile of notebooks.

Now comes the part that decides whether any of it counts: you have to write it down so a stranger can run with it.

That is not a metaphor. The whole machinery of science — peer review, replication, citation, the slow accumulation of trustworthy knowledge — runs on the paper. A result that lives only in your head, or only in a notebook only you can read, is not yet knowledge. It is a private belief. The paper is the artifact that turns your private belief into a public claim that other people can check, attack, build on, or refute. Habakkuk’s instruction to the prophet is the instruction to every researcher: write the vision, make it plain, so he may run who reads it. The reader has to be able to run — to take your method and move.

This is a drill week. There is no separate project file. The thing you build this week — the Draft Paper (10% of your grade) — is the lab in the exercises. You will take every deliverable you have produced and assemble it into a single conference-style manuscript in a real publication template: ACM acmart or IEEE IEEEtran, the same templates you started in Week 1 (see Appendix A). By Friday you have a draft that a peer can review in Week 15.

The AI thread runs through this chapter from both sides. As a tool you wield: a 2026 LLM is a genuinely useful drafting partner — it will tighten a clunky sentence, propose an outline, suggest a clearer figure caption. As a workload you must govern: the single most dangerous thing an LLM does to a paper is hallucinate citations. Walters & Wilder (2023) found 55% of GPT-3.5 citations and 18% of GPT-4 citations were entirely fabricated, and among the real ones, 43% / 24% had substantive errors. A fabricated reference in your bibliography is not a typo. At ICCV it gets your paper rejected without review; arXiv (May 2026) has threatened a one-year submission ban for “incontrovertible evidence” of unchecked LLM content — hallucinated references, leftover meta-comments, unreplaced placeholder data. The human owns every word, every figure, and every citation. The model drafts; you verify; you are accountable.

So the apologetic question this week is sharp and practical: How do you write the vision and make it plain? What does it mean to communicate truth so faithfully that another person can stand on it? We will take that seriously in §14.9 and §14.y.

Coach’s Note — A paper is not a transcript of what you did. It is an argument about what your results mean, organized so the reader believes you for the right reasons. You did the experiments in one order; you will present them in another. That is not dishonesty — it is rhetoric in the old, honorable sense. The line you may never cross is changing what is true. You reorder for clarity; you never reorder to deceive.

14.1 — The Standard Structure: IMRaD, Expanded for CS

Empirical CS papers follow a structure so consistent that reviewers read them in a fixed order and get annoyed when you deviate. The backbone is IMRaD — Introduction, Methods, Results, and Discussion — expanded for our field:

The hard discipline is the Methods / Results / Discussion firewall. Methods is what you did (written so another team can rerun it). Results is what you found (no spin — the numbers, the tables, the figures, reported neutrally). Discussion is what it means (interpretation, comparison to prior work, the honest limitations). New researchers smear these together — they interpret in Results, or sneak results into Methods. Keep them clean. A reviewer who can find the firewall trusts you more.

Notice what this structure encodes: you are forced to separate observation from interpretation. That separation is the heart of research integrity. You report Exodus 20:16 — “you shall not bear false witness” — in the Results section by reporting what you saw, not what you hoped. You make meaning in the Discussion. The structure itself is a guardrail against self-deception.

Coach’s Note — Write the sections in the order that is easiest, not the order they appear. Most experienced researchers write Methods first (you already know what you did), then Results (the numbers are done), then Discussion, then Related Work, then Introduction, and the Abstract dead last. The abstract is a promise; you can only make it once you know what you delivered.

You are not starting from a blank page, and this is the payoff of keeping a portfolio all semester. Your Methods section is your experimental-design document (Chapter 7) plus your reproducibility setup (Chapter 9), rewritten for a reader instead of for yourself. Your Results section is your raw results (Chapter 11) and your analysis (Chapter 12), with the dead ends pruned and the figures cleaned. Your Related Work is your comparison matrix (Chapter 4) turned into prose. Your Discussion is Chapter 13. The draft this week is less an act of writing than an act of assembly and translation — which is exactly why a practicum that produced real artifacts every week converges on a paper, while a class that only lectured on method would now face a blank document and a panic.

A skeleton that already has these headings, the firewall comments, and placeholder counts is in code/paper-skeleton.tex. Start from it.

14.2 — The Abstract: Five Moves in Two Hundred Words

The abstract is the most-read and worst-written part of any paper. Most people will read only this. It must stand completely alone: no undefined acronyms, no “see Section 4,” no citations unless unavoidable.

The most reliable scaffold is Kent Beck’s four-sentence abstract (OOPSLA 1993), which we extend to five moves for empirical work:

1. Context — the world the problem lives in. (“Modern code review increasingly relies on LLM agents…”)
2. Gap / problem — what is missing or wrong. (“…but their reliability on real defects is unmeasured against human reviewers.”)
3. Approach / contribution — the startling sentence: what you did that is new. (“We evaluate three LLM review agents on 500 SWE-bench Verified tasks against a human baseline.”)
4. Result — the concrete finding, with a number. (“Agents catch 71% of seeded defects but miss 6 of 8 logic errors humans flag.”)
5. Significance / implication — why a reader should care. (“This bounds where LLM review can be trusted unsupervised and where it cannot.”)

Beck’s insight: the third sentence — the contribution — is the one you write first and sweat most, because it is the only sentence that makes your paper different from every other paper. The other four exist to make that one land.

Coach’s Note — Put a number in the abstract. “Our method improves performance” is invisible. “Our method reduces 95th-percentile tail latency by 31% (95% CI [24%, 38%])” is a claim a reader can grade. Vague abstracts read as papers with nothing to say.

The Paper Structure Builder widget (§14.x) enforces these five moves: it will not let you mark the abstract done until each move is present.

14.3 — The Introduction: Swales’ CARS in Three Moves

The introduction is where you earn the right to the reader’s attention. The cleanest model is John Swales’ CARS — Create A Research Space (1990), three rhetorical moves:

1. Establish the territory. Why this area matters; what is generally known. (Brief. The reader is not stupid.)
2. Establish the niche. The gap — Move 2. “However, prior work assumes X / has not measured Y / breaks under Z.” This is the sentence your whole paper exists to fill. Back it with a cited absence, not a hunch.
3. Occupy the niche. What you do about it: your approach, and an explicit contributions list.

End the introduction with a bulleted contributions list. Reviewers look for it; many score it directly:

This paper makes the following contributions:
  C1. We construct the first benchmark of N real defects for LLM code review.
  C2. We measure three 2026 review agents against a human baseline (§4).
  C3. We characterize the failure modes (logic vs. style) that bound safe use.

If you cannot write three concrete, checkable contributions, you do not yet have a paper — you have an experiment that needs a sharper claim. Go back to your research question (Chapter 6) before you keep writing.

Coach’s Note — A reviewer decides whether to like your paper in the first page. The introduction is not a warm-up. By the bottom of page one they should know your gap, your contribution, and your headline result. Do not save the punchline.

14.4 — Related Work: Position, Don’t Summarize

Related Work is not your annotated bibliography from Chapter 3, and it is not a list of summaries. It is an argument that organizes prior work into camps and positions you against them. The reader should finish the section knowing exactly what you do that no one else has.

Pull this straight from your comparison matrix (Chapter 4) — rows are papers, columns are problem / method / dataset / metric / result / limitation. The empty cell, the unaddressed column, is your gap, and Related Work is where you name it out loud. Group by approach, not by paper: “One line of work uses X [3, 7, 12]; these achieve A but assume B, which fails in our setting.” Then the killer sentence: “In contrast, we…”

The integrity rule here is severe: you must represent prior work fairly. It is tempting to make a competitor look weaker than it is so your contribution shines. Do not. 2 Corinthians 13:1 — “every charge must be established by the evidence of two or three witnesses.” Cite accurately, quote fairly, and never knock down a strawman. A reviewer who has read [12] will notice, and your credibility dies on the spot.

Coach’s Note — If you are writing for arXiv eventually, know this: as of Oct 31, 2025, arXiv’s CS category will not accept review/survey/position papers unless they are already peer-review-accepted at a journal or conference (workshop review is explicitly not enough). Your Related Work section is fine — but a standalone literature survey is not arXiv-postable from a practicum. Plan for a workshop instead (Chapter 16).

14.5 — Methods and Results: The Firewall, in Practice

Methods is the section a replicator reads with your repo open. Write it so a competent stranger could rerun your study without emailing you. From your reproducibility work (Chapter 9), surface: the data and its splits, baselines and ablations, the metric and why it operationalizes your construct, the number of seeds/runs, hardware, and software versions. Do not narrate your debugging journey — describe the final, clean procedure.

If an LLM was an instrument in your method (a judge, a generator, an agent), Methods is where you say so precisely: the model name and version, the date, the temperature, and the exact prompt (in an appendix if long). Closed models drift, which silently breaks reproducibility — so the version-and-date stamp is not optional. NeurIPS 2026’s policy is that LLM use is described in the experimental-setup section only when it is an important, original, or non-standard part of the method — which an LLM-as-instrument absolutely is.

Results reports what you found and nothing about what it means. Tables and figures carry the load; prose points to them. Report effect sizes and confidence intervals, not just p-values (Chapter 12; ASA 2019). If your benchmark numbers come from multiple seeds — and they should (Bouthillier et al. 2021) — report the spread, not a single lucky run. A single-run “we beat the baseline” is statistically empty.

Order matters inside Results too. Lead with the headline result that answers your research question, then the supporting breakdowns, then the ablations that isolate why it works. Each subsection should map to a contribution from your introduction (C1, C2, C3): the reader who read your contributions list should be able to find the evidence for each one without hunting. If a result does not serve a stated contribution, ask whether it belongs in the paper at all — or whether you under-claimed in the introduction and should add a contribution.

Here is the part where you will be tempted, and where you must not yield. You ran experiments and some of them “didn’t work.” Those runs go in. The integrity of the record (Chapter 11) is the integrity of the paper. You report what you saw, not what you wished. Dropping the inconvenient runs is falsification — one leg of the FFP triad (fabrication, falsification, plagiarism) that defines research misconduct under U.S. federal policy (42 CFR Part 93).

Coach’s Note — A clean test for the firewall: read each sentence in Results and ask, “Could a hostile reviewer dispute this without re-running my experiment?” If yes — if it requires interpretation — it belongs in Discussion. “Method A scored 0.81, Method B scored 0.74” is a Result. “Method A is better suited to production” is a Discussion claim.

14.6 — Figures, Tables, and Captions That Stand Alone

A reviewer skims your figures before reading a word. If your figures and captions cannot carry the argument by themselves, you lose readers who would have agreed with you.

Rules that hold across venues:

• The caption is self-contained. A reader who jumps to the figure understands it without the body text. Tables get captions above; figures get captions below (ACM/IEEE convention). State what is shown, the units, the sample size, and what the reader should conclude.
• Tables for exact numbers; figures for trends and shapes. Do not screenshot a table as an image.
• Every figure is referenced in the text, in order, before it appears: “Figure 2 shows…”. An unreferenced figure is a reviewer red flag.
• Show uncertainty. Error bars or shaded CI bands, with the caption stating what they represent (SD? SE? 95% CI?). A bar chart with no error bars, after Chapter 12, is a confession.
• Honest axes. No truncated y-axis that turns a 2% difference into a cliff. A misleading figure is a “false balance” (Prov 11:1) rendered in pixels.
• Vector, not raster. Export PDF/SVG so figures stay crisp in the two-column PDF. matplotlib.savefig("fig2.pdf"), not .png.

Coach’s Note — Write the caption as if it will be tweeted out of context, because it will be. Conference attendees photograph one slide; readers screenshot one figure. Make that one unit true and complete on its own.

14.7 — The Templates: acmart, IEEEtran, and Overleaf

You started this in Week 1; now you commit to a real venue’s template. The two that cover most of applied CS are pre-loaded on Overleaf, so no local LaTeX install is required (Appendix A).

(Versions confirmed via search summaries as of mid-2026; re-check the literal version string on the venue’s current template page before you submit — these tick over.)

Pick the template that matches the venue you are aiming at (you will lock the venue in Chapter 16). For ACM, the typical submission flow is single-column manuscript,review; for double-blind venues, add anonymous and scrub identifying info now — it is far easier than retrofitting later. Manage references with BibTeX: keep a references.bib, cite with \cite{key}, and — this is the rule that matters — every entry must be verified against a real index (dblp, Semantic Scholar, or the publisher). A starter references.bib with three correctly-formatted, real entries is in code/references.bib; a pre-flight checklist before you compile is in code/repro-checklist.txt.

% acmart submission preamble (single-column, anonymous for double-blind review)
\documentclass[manuscript,review,anonymous]{acmart}
\begin{document}
\title{Bounding LLM Code-Review Reliability on Real Defects}
\begin{abstract} ... \end{abstract}  % five moves, §14.2
\maketitle
\input{sections/intro}     % CARS, §14.3
\input{sections/related}   % position, §14.4
\input{sections/method}    % the firewall, §14.5
\input{sections/results}
\input{sections/discussion}
\bibliographystyle{ACM-Reference-Format}
\bibliography{references}   % verified BibTeX, §14.7
\end{document}

Coach’s Note — Overleaf has both templates under “Templates → search ‘acmart’ / ‘IEEEtran’.” Compile early and compile often. A LaTeX error the night before a deadline, in a document you have never successfully built, is a self-inflicted wound. Build a one-page skeleton in week one of writing, confirm it compiles, then fill it.

14.8 — AI as a Drafting Partner — and Where It Stops

A 2026 LLM is the best prose-editor a graduate student has ever had at hand, and you should use it. But you must know exactly which jobs it is good at, which jobs are yours alone, and where the line between them sits.

Use it for these. Tightening a tangled sentence you already wrote. Proposing three alternative phrasings for an awkward transition. Suggesting an outline you then rearrange. Catching the passive voice you cannot see anymore after the fourth read. Drafting a first-pass figure caption you will rewrite. Checking whether your abstract actually hits all five moves. These are assistive uses — closer to a very good grammar checker — and at most venues they need no disclosure (ACM and IEEE both exempt grammar/editing aids).

Never delegate these. The claim. The contribution. The decision about what the data mean. The judgment that a competitor was represented fairly. And — above all — the citations. The moment an LLM supplies content rather than polish — a generated paragraph of related work, an invented statistic, a reference — you have crossed from assistive to generative, and three things become true at once: (1) you must verify every word against reality, (2) you may owe a disclosure (the rule is venue-specific; see Appendix C), and (3) you, the human, remain fully accountable regardless.

The failure modes are concrete and current. A 2026 study of long-context Q&A over 172 billion tokens (arXiv:2603.08274) found that hallucination rises with context length — exactly when you paste your whole draft in and ask for “a polished related-work section,” the model is most likely to invent. Fluent and unfaithful are not opposites for an LLM; they are frequent companions. The output reads beautifully and asserts things that are not so.

Coach’s Note — Here is the spine rule for this whole book, applied to writing: the human stays in the loop where the judgment lives; AI accelerates, the human decides, verifies, and is accountable. For a paper, the judgment lives in the claim and the citation. Let the model carry your sentences; never let it carry your conscience.

14.9 — Citation Hygiene and the Discipline of Honest Attribution

This is the section that separates a researcher from a content generator, so slow down.

A citation is a promise: this claim is supported by that source, and I have read enough of that source to know it. Break the promise and you break the chain of trust that all of science runs on. There are three ways to break it, in ascending order of seriousness:

1. Sloppy metadata — wrong year, wrong venue, mangled author list. Bhattacharyya (2023) found 87% of LLM-generated citations to real works had at least one metadata error. Annoying; fixable; still your fault.
2. Citing what you did not read — pulling a reference from another paper’s reference list without opening it, and inheriting their misreading. Common, lazy, and how errors propagate for decades.
3. Hallucinated citations — a reference to a paper that does not exist. This is the LLM failure mode. 55% of GPT-3.5 and 18% of GPT-4 citations were entirely fabricated (Walters & Wilder 2023). Grounded/RAG tools (Elicit, Consensus, scite) hallucinate less because they cite from a real index — but you still verify every one.

The rule is absolute and simple: verify every reference against a real index before it enters references.bib. Open dblp or Semantic Scholar, find the paper, confirm the authors/venue/year, and copy their BibTeX. If you cannot find it, it does not go in — no matter how confidently a chatbot produced it.

This is where the apologetic question lands. Write the vision; make it plain, so he may run who reads it (Hab. 2:2). The reader runs on your citations. When you cite a source you never read, or a source that does not exist, you have made the vision unrunnable — you have laid a path that collapses under the foot of the next researcher. Honest attribution is a form of love for that future stranger. The LCMS confessional tradition names the deeper logic plainly in its explanation of the Eighth Commandment (“You shall not bear false witness”): we are not merely to avoid lying, but to “defend [our neighbor], speak well of him, and explain everything in the kindest way.” Citing a rival’s work fairly — representing it at its strongest, not its weakest — is exactly that command applied to scholarship. The temptation to inflate your contribution by deflating theirs is, at bottom, a temptation to bear false witness.

And there is a humility in it. The scientist does not stand alone; she stands, as Bernard of Chartres put it, on the shoulders of giants. Deuteronomy 32:7 — “Remember the days of old; consider the years of many generations.” Every citation is an act of remembering rightly. Research is not the heroic solo discovery of myth; it is a slow, communal, multi-generational searching-out of a world we did not make and do not fully comprehend — “the glory of kings is to search things out” (Prov. 25:2). To write a paper well is to take your honest, bounded place in that long conversation, and to leave the record true for whoever comes next. Truth-telling in a bibliography is a small thing that rests on a very large one.

The venues encode this. An LLM can never be an author at any major CS venue — ACM, IEEE, NeurIPS, ICLR, ICML, ACL, CVPR, arXiv all agree — because authorship requires accountability, and a tool cannot be accountable (the ICMJE criterion the model fails is final approval + accountability). “The model did it” is never a defense. The human owns every claim. (Where and how to disclose legitimate AI assistance is venue-specific; see Appendix C for the ACM/IEEE/NeurIPS/arXiv rules, and disclose per your target venue.)

Coach’s Note — Run this drill before every submission: take five references from your .bib — especially any you first encountered through an AI tool — and find each one in dblp by hand. If even one will not resolve, audit the whole bibliography. It is the cheapest insurance in research, and it has saved careers.

14.x — Interactive Lab: Paper Structure Builder

Open the Paper Structure Builder embedded on this page, below the chapter. It is the warm-up for this week’s graded lab.

Work through it in two passes:

1. Outline the whole paper, section by section. Drag your own deliverables into each slot — your gap statement into the Introduction’s niche, your comparison matrix into Related Work, your design doc into Methods, your results tables into Results, your interpretation into Discussion. As you fill slots, the builder flags the firewall violations: an interpretation sentence sitting in Results, a result that leaked into Methods. Fix them in the tool before you write them in LaTeX.
2. Build the abstract. The abstract builder enforces Beck’s five moves — context → gap → approach → result → significance (§14.2). It will not let you mark the abstract complete until all five are present, and it will warn you if your result move has no number in it.

The widget teaches the shape of the paper so that when you sit down in Overleaf you are filling a structure you already understand, not staring at a blank \documentclass. Treat its output as the outline you hand to your peer reviewer in Chapter 15.

14.y — A Theological Aside: Making the Vision Plain

We touched this in §14.9; here we sit with it, because the apologetic question of this week — how do you write the vision and make it plain? — is not decoration on a writing chapter. It is the writing chapter.

Habakkuk is given a vision and a command in the same breath: not merely to receive truth but to transcribe it so legibly “that he may run who reads it.” The point of plainness is motion — the reader must be able to act on what you wrote. A paper so tangled that no one can reproduce it has failed Habakkuk’s test even if every word in it is true. Clarity is not a courtesy you extend to the reader after the science is done; clarity is part of the science, because an unrunnable result is not yet knowledge.

This reframes the work of writing. The all-nighter spent untangling a muddled Methods section, the third rewrite of an abstract, the hour spent verifying one citation — these are not the unglamorous tail end of “real” research. They are the act by which private finding becomes public truth. The researcher writes plainly for the same reason the prophet does: because the truth is not yours to hoard, and the next person has to be able to run with it.

It also bounds your ambition honestly. You will be tempted to write the vision you wish you had — the cleaner result, the bolder claim, the rival diminished. Habakkuk’s command is to write the vision, the one actually given, plainly. The discipline of the Results-Discussion firewall, the insistence on effect sizes over cherry-picked p-values, the verified bibliography — these are spiritual disciplines as much as methodological ones. They are how you tell the truth about a world you are searching out but did not author. “Test everything; hold fast what is good” (1 Thess. 5:21) is the reviewer’s creed and the writer’s: hold fast to what the data actually support, and write that plainly.

There is one more thing worth naming, because the academy can make you forget it. The paper has your name on it, and a published result can buy a degree, a job, a reputation. The pull toward overclaiming is not abstract; it is the very concrete pull of self. But the confessional tradition you stand in says the work was never finally for that. “Whatever you do, work heartily, as for the Lord and not for men” (Col. 3:23) — the verse that ends this whole course in Chapter 16 — reframes the audience of every sentence you write. The reader you are ultimately writing plainly for is not the program committee. It is the next searcher-out of truth, and behind that one, the God who concealed the thing for kings to find (Prov. 25:2). Knowing that does not make the writing easier. It makes the temptation to lie smaller, because there is no one left to impress and no result worth purchasing with a false witness.

14.z — Common Pitfalls

Pitfall: Interpreting in the Results section. Example: “Method A scored 0.81, demonstrating its superiority for production deployment.” Fix: Stop at the number in Results. Move “demonstrating its superiority…” to Discussion, where a reviewer expects — and can challenge — interpretation.

Pitfall: Dropping the runs that “didn’t work.” Example: Three of ten seeds underperformed the baseline, so you report the mean of the seven good ones. Fix: Report all ten. Falsification (omitting inconvenient data) is research misconduct under 42 CFR Part 93, not editorial license. Report the full spread; discuss the variance honestly (Bouthillier et al. 2021).

Pitfall: Trusting an LLM’s bibliography. Example: You asked a chatbot for five references, pasted them into references.bib, and one of them does not exist. Fix: Verify every entry against dblp / Semantic Scholar / the publisher before it enters the file. 18–55% fabrication is not a tail risk; it is the base rate (Walters & Wilder 2023). Copy BibTeX from the real index, never from the chatbot.

Pitfall: A vague, number-free abstract. Example: “We propose a novel method and show it improves performance significantly.” Fix: Put the headline number with its uncertainty in the abstract: “improves recall by 14 points (95% CI [9, 19]) over the strongest baseline.” Invisible claims read as no claim.

Pitfall: Related Work as a list of summaries. Example: “Smith (2024) did X. Jones (2025) did Y. Lee (2026) did Z.” Fix: Organize by approach and position yourself: “Retrieval-based methods [Smith; Jones] achieve A but assume B; in contrast, we…”. Pull the structure from your comparison matrix (Chapter 4).

Pitfall: A figure that does not stand alone. Example: A plot captioned “Figure 3: Results,” with no units, no sample size, no statement of what the bands mean. Fix: Write the caption as if it will be seen out of context: what is shown, units, n, what the error bands represent (SD/SE/CI), and the one thing the reader should conclude.

Pitfall: Misrepresenting prior work to make yourself look better. Example: Describing a competing method’s known limitation as its central claim, so your contribution looks larger. Fix: Cite fairly — “every charge established by two or three witnesses” (2 Cor. 13:1). A reviewer who has read the source will catch the strawman and stop trusting you.

14.(z+1) — Reps

The week’s work is in the exercises, and this week the reps are the graded lab: you assemble the full Draft Paper in your ACM/IEEE template. A preview:

• Rep 1 — Stand up the Overleaf project from the correct template and confirm it compiles.
• Rep 2 — Write the abstract via Beck’s five moves; paste it into the widget to confirm all five are present.
• Rep 3 — Draft the introduction with CARS’s three moves and an explicit contributions list.
• Rep 4 — Convert your comparison matrix into a positioned Related Work section.
• Rep 9 — Run the citation hygiene audit: verify every references.bib entry against dblp or Semantic Scholar.

There is an on-page Check Your Reps quiz below the chapter — five questions on structure, the abstract’s five moves, the Methods/Results firewall, citation hygiene, and the templates. Take it before you start writing; it is the fastest way to catch a misconception while it is still cheap.

14.(z+2) — This Week’s Deliverable

No separate project file this week — this is a drill week. The lab in the exercises is the graded Draft Paper (10% of your final grade): a complete conference-style manuscript in acmart or IEEEtran, with every section present (abstract through future work), real verified citations, figures and tables that stand alone, and the Methods/Results/Discussion firewall intact. It does not have to be polished — it has to be complete and honest, every section drafted, so that a peer can review the whole arc.

That draft goes straight into Chapter 15 — Peer Review and Revision, where two classmates will review it under double-blind rules and you will review theirs. Write the draft this week as though a stranger will have to run with it — because next week, one will.

14.(z+3) — Coach’s Final Word

You have been doing research for thirteen weeks. This week you become an author. The difference is accountability: a notebook answers to you, a paper answers to everyone who reads it.

Write the sections in the easy order, build the abstract last, keep the firewall clean, and verify every single citation by hand. Resist the two great temptations — to overclaim what your data show, and to underclaim what your rivals did. Both are forms of false witness; both a good reviewer will catch; both cost you the only currency you have, which is trust.

Make the vision plain. Then hand it over and let someone run with it.

See you on Monday.

Up next: the graded lab in the exercises — assemble the full Draft Paper — then Chapter 15 — Peer Review and Revision. Reference material: Appendix A (research environment), Appendix B (researcher’s toolkit), Appendix C (using AI responsibly), Appendix D (glossary). Previous: Chapter 13 — Discussion and Interpretation.